One thing before anything else, because it changes how you read the rest: the whole desk is open source, which means the code is public and free for anyone to download and change, and the link sits at the end of this piece, so you can take mine and start from it, or follow how every tab got built and make your own, and I am explaining the whole process here so that either route works for you.

Anyone who takes markets seriously ends up with the same setup, and it grows on you without you noticing it. You have Finviz, Koyfin or Screener.in for the screening, OpenInsider for the insider filings, Capitol Trades for what Congress is buying, WhaleWisdom for the 13F filings, Quartr for the earnings call transcripts, OptionCharts for the options data, FRED and Trading Economics for the macro, Portfolio Visualizer for the backtests, Stock Analysis or Tijori for a clean set of financials, and a forum like ValuePickr for the discussion, because you tried each of them at some point and each one did one part of the job well enough to keep. Five or six tabs, seven or eight on a heavy day, and most of them are on a subscription.

Now, paying for them is not the problem. Some of those tools are definitely good at the one thing they do, and I am not going to tell you to cancel anything. The problem is that none of these tools talk to each other, which means you are the one doing the talking. You take a margin number out of the screener, you go and check it against the filing, you put it into your notes, and then you go to a fourth window to set the alert, and you do that ten times a day, twenty times a day, and after a while it feels normal, because every tool in that stack is automated and the only part that is not automated is the bit in between, which is you.

That tab stack was my setup for years, honestly, and I did not think there was another way to do it, because the other way was supposed to belong to institutions. A fund has a Bloomberg terminal or FactSet on every desk, one screen where the positions, the filings, the estimates, the ownership data and the alerts all sit together, and for a long time the assumption was that an individual investor could not have anything like that, because those systems are built for institutions and sold to institutions. That assumption is now wrong. What you are looking at above is a small Bloomberg for one desk, a mini FactSet if you like, with three portfolios across two markets in one window, and every piece of it is a file on my machine. This edition is the whole thing, start to finish, including how each of those tabs got replaced, and you can build it in your own market with your own broker, which is the part I want to get right for you.

A Claude Code desk has five layers

Before I walk you through the tabs, I want to give you the map, because without the map the twelve tabs look like twelve more tabs, and that is exactly the thing we are trying to get away from.

The way I think about it, the desk has five layers. There is a memory, there are workflows, there is a clock, there is a screen, and there is the data underneath, and the thing all five have in common is that none of them is a service I log into. Every one of them is a file, or a folder of files, sitting in one folder on my Mac, and I am going to explain each one properly, so if you have never read the earlier editions you are not missing anything.

The memory is a vault

If you have not used Obsidian, it is a free note-taking app, and the only thing you need to know about it is that a vault is just a folder of plain text files on your computer, written in markdown, which is plain text with a few symbols for headings and links. There is no database and no account behind it, and inside my folder there are four things that matter: a file called CLAUDE.md at the top, which is the standing instruction sheet, how I want things done, where things live, what never goes out; a folder called wiki, where every company I have looked at, every filing note and every rule I have written for myself sits as a note; a folder called raw, the inbox, where anything I clip or drop in lands first; and a folder called output, where everything the desk produces goes.

Around that folder sits a small set of plugins, one that clips any web page into the vault, one that keeps the charts inside the clip, one that opens the agent in a terminal next to your notes, one that opens the desk and any web page inside Obsidian, one that puts the agent in a chat window if you never want a terminal, and a browser the agent drives itself, and every one of them gets installed in the walkthrough near the end, with screenshots, so you do not need the names yet.

The agent that reads all of that is, in my case, Claude Code. If you have not seen it, it is Anthropic’s terminal tool, which sounds technical and is not. You open it on a folder, you type in plain English, and it can read the files in that folder, write new ones, and run programs on your machine. That is the whole difference between it and a chat window: a chat window gives you an answer, and this thing does the work and leaves the result as a file. I open it on the vault, so the vault is what it remembers. Kimi Code, Grok Build and Codex are the same kind of tool from other companies, and I use Kimi Code and Grok Build inside Obsidian as well, sometimes on the same day, and if a terminal is the thing that has kept you away from all of this there is a desktop-app route, which sits in the install section near the end with everything else.

Nothing in this setup is married to Claude. The reason everything is plain text is that any model can read plain text, and the reason the rulebook is one file is that any agent can be told to read it. Claude Code reads CLAUDE.md on its own. The other agents look for a file called agents.md, and my instruction to them is simply to read CLAUDE.md, so there is one rulebook and nothing is locked inside anybody’s product. That was the argument I made in July when I rebuilt the setup for the second time, and it is still the argument.

The workflows are skills

This is the layer the earlier setups never had. A workflow here is a skill, which is a job written down once in a plain text file, the sources, the steps, the checks, the shape of the output, so the agent runs it by name instead of me typing the same instructions every time. And here is the thing about the desk that took me a while to see: most of the twelve tabs are skills too. The Funds tab is a skill that goes to EDGAR and pulls the filings, Capitol is one that reads the Congress disclosures, Short reads FINRA, Macro reads FRED, Flow reads the options chains, and you do not fire any of them, you click the tab and the job runs itself and refreshes on its own, which, honestly, is the coolest part of the whole thing. The few I still fire by name come back in the section on what runs on its own.

The clock runs without me

The clock is everything that runs without me asking, the desk’s own refresh loops, a set of alert rules that ping my Mac when one fires, and the daily rituals I fire with one phrase, and it gets its own section later, including the way to keep all of it running when your laptop is shut.

The screen is the dashboard

The screen is the dashboard you saw at the top, one small program, written by the agent, that shows twelve tabs on my own machine with nothing else to install and nothing to host, and it sits inside Obsidian in a tab like any page, which is why the desk and the notes are in one window. The fifth layer, the data underneath, is the next section, because it is the question everyone asks first and it decides whether this works in your market, and the tour of the tabs comes right after it.

If you want to see how the older versions of this setup looked, the May edition covered the first install and the July edition covered the rebuild, and both are linked at the bottom of this piece, though nothing here depends on them. And one more thing, because a few of you will want to watch this rather than read it: I am recording the whole setup running live, as a video walkthrough, and it goes out as a companion to this piece over the next couple of weeks.

A small note, and the only sponsor here is me. I keep getting messages asking how I set up Claude Code for stock research, beyond what the editions can teach. So I am opening a small consulting program: I set your system up with you, then support sessions over the following months as your real problems show up. I am doing a lot of other things, so seats are very few. If that is you, the waitlist is on the website or reply to this email.

Where the desk’s data comes from

Three places, and they do three different jobs, so let me take them one at a time.

Source one, your broker

It is the only source in the world that knows what you actually own, at what price, with how much margin left. My accounts come through the broker’s own API, which is the door a broker opens so a program can read your account, and the desk reads holdings, positions, funds and margin straight from it. Most brokers can be connected now, through an official API or a connector built for AI agents: Interactive Brokers almost everywhere, Schwab, Alpaca and Robinhood in the US, with Robinhood’s new connection built for AI agents specifically, read access to your accounts and orders kept inside a separate account you fund for it, Trading 212 and Saxo in Europe and the UK, Zerodha, ICICI and the other large brokers in India, Questrade in Canada. Where a broker offers no connection, a small text file with your positions does the job and the desk prices it live. You connect whichever broker you have.

Source two, an optional feed

This is the one everyone assumes is the whole game, and it is not. A feed gives you breadth: screening across a few thousand names, years of price history, peers, analyst estimates, and the financial statements already parsed into rows. There are many providers, and you check the ones that cover your market. I use Financial Modeling Prep because I only need it for US names and its US data is good, on its cheapest plan, and the plans above it add Canada, the UK and then global coverage. The reason I pay for a feed at all is that I look at financials and ratios on different names all day, and I do not want the agent spending one minute, two minutes, five minutes fetching a new company every single time. For names the feed does not cover, that is exactly what happens, and it works: the agent opens a real browser through Playwright, which is a browser it drives itself, and reads the numbers off the exchange site or a site like Screener.in, or uses its own fetching tools, and if you want a service built for reading web pages in bulk, Firecrawl has a free monthly quota and a paid plan above it. So the feed is a convenience, not a requirement. If you do not want one, skip to source three, and nothing breaks.

Source three, the filing itself

Every market has a public disclosure system. In the US it is EDGAR, the SEC’s filing database, where every annual report, the 10-K, sits the day it is filed, and everywhere else there is an equivalent, the exchange portal, the regulator’s site, the annual report on the company’s own page. It is free, it is the primary record, and until recently it was expensive to use, because reading a two-hundred-page filing takes an afternoon, which is why everyone bought the feed instead, and now the agent reads it.

Let me show you, because this is the button every one of you can click. Here is the prompt, exactly as I ran it the first time:

Pull S&P Global's 10-K filings from EDGAR. Give me five years of revenue, operating profit, net income, diluted EPS, operating cash flow, capex, buybacks, year-end cash and diluted shares. Put the accession number of the filing each figure came from next to it.

What the run did, in plain words, is this. It asked the SEC’s filing index for S&P Global’s annual reports, it took each number from the tagged data the company itself files with the report, and it wrote the accession number next to each one. The accession number is the filing’s serial number on EDGAR, so anyone can open the exact document a figure came from, and I did exactly that for the 2025 and 2024 columns, checked them against the statement pages in the 10-K, and they match. The whole thing took a few seconds, with no subscription and no login.

The second time, just the name

Now, the part that matters more than the prompt, and I want to be careful here, because those nine items are mine, not a standard. I typed that prompt in full once. The second time, I typed “analyse S&P Global”, and the agent did all of it on its own, because the job was now written down in the setup. And the job is whatever your framework is. If you read a company top-down from the industry, or bottom-up from the unit economics, if there are twelve ratios you always want first or three sections you read before anything else, you tell the agent once, you tell it to remember, and from then on “analyse” means your method, on every company you hand it. It does not fit you to a fixed list. It fits itself to you. That is the difference between this and a tool like Koyfin or Finviz. Those give you their features and you fit your process to them. Here the features get built around your process, and next week it adds whatever you ask for, which is why my setup is completely different from yours and both are right. That is the whole point.

And it keeps getting better, which nobody tells you. It makes mistakes. It misses something you needed, or it gives you something extra you did not ask for, and sometimes that extra thing is the useful one. You give it the feedback, it fixes the instruction, and it does not make that mistake again, and this goes on every day, so my setup today is better than it was a month ago and a month from now it will be better again, and every model upgrade makes the same folder sharper. The closest thing I can compare it to is having an analyst of your own who thinks like you and works like you, gets something wrong now and then, and actually improves when you tell them.

Say you are in the UK, or Dubai, or anywhere else. You do not even have to tell it where the filings live. The first time, it works out the portal on its own, the exchange, the regulator, the company’s own site, it pulls your list, and the reference is whatever that market uses, a filing link, a page number, instead of an accession number. Once it has done that correctly once, you tell it to remember, and that is the standing method for that market.

So the rule I run the desk on is that the broker knows your money, the feed, if you have one, tells you where to look, and the filing decides what is true, and I do not let the first two overrule the third.

The Funds tab runs the same way

The desk itself is built on that rule, not just the demo, and the Funds tab is the cleanest example, because it does not touch the paid feed at all. Every card on it is a 13F filing, the quarterly form where large funds disclose their US holdings, read straight from EDGAR, the same free public record as the 10-K above. And the way it got built is what I want you to take away. Once the dashboard existed, I told the agent, in plain English, something very close to this: here is the list of funds I follow, go to EDGAR, pull each one’s latest 13F, show me the top holdings with the weights, compare it with the previous quarter and show me what is new, what was sold, what was added to and what was trimmed, and pick up the new filing when it lands. That was the instruction. It built the tab, and the list of funds sits in a small file I can edit whenever I want to follow someone new.

Take one card so you can see what it actually holds. Druckenmiller’s Duquesne filed its quarter ending June 30, 2026 on August 14, 89 positions, about 5.2 billion dollars, and the card reads the two most recent filings against each other. New this quarter: Fox, CDW, Alphabet, Bitdeer, Delta, Tesla. Exited: Broadcom, Option Care Health, Figure Technology, Restaurant Brands, Lattice Semiconductor, Humana. Added to, and read these percentages carefully, because they are how much bigger the fund’s position got, not how the stock moved: the iShares position is 318 percent larger than last quarter, the Amazon position 307 percent larger, United Airlines 203 percent, PureCycle 178 percent, Seagate 141 percent. Trimmed, same measure: the Alcoa position cut by 88 percent, Roku by 75 percent, Teva by 74 percent, Arm by 74 percent. Under that sit the top holdings with the fund’s weight and what share of each company it owns. Every one of those lines came out of two public filings and a comparison, and it is there for every fund on the list, every quarter, without opening WhaleWisdom again. It is not a view on any of those names. It is what the filing says.

What the twelve tabs show

So this is the tour, tab by tab, and I am keeping each one to what it shows and what it reads, because the desk is the point here and not any one number on any one screen. Two of the tabs you have already met, the first desk at the very top of this piece and the Funds tab in the last section, so I will not walk them again.

The two desks, priced live

The first desk you saw in the cold open, both accounts with the futures positions, the options tape on the index and the margin bars. The second desk is our US account, and it has three parts, the positions at the top, priced live, the earnings countdown for every name on the book and the watchlist, and a pulse of the whole US market, and further down the same page an insider tape off the Form 4 filings, which are the forms insiders file when they buy or sell their own stock, with the cluster buys under it, where two or more insiders bought the same name in the same window, which is the one insider pattern I actually act on. Any ticker on any of these pages is one click away from its own research page, which is the next section.

Three watchlists, one grid each

Watch·IN reads the broker with live ticks, Watch·US reads the feed, and Global reads Yahoo’s free quotes with no key at all, which is how one grid holds indices, stocks from seven exchanges, the currency pairs and the commodities. The quotes cycle through each list about once a minute, the price flashes when it ticks, and a badge comes up when a name is at the day’s high or low. You add a name by typing its symbol at the top, and each list is one small file of symbols, so adding a market is adding a line to it.

The intelligence tabs read the public record

The other three intelligence tabs work the way Funds does, a public dataset read on a schedule and shown for the names I own or watch.

Flow reads the options chain, the full list of puts and calls trading on a name, for every US name on the book and the watchlist, from the options exchange’s free delayed data at CBOE, and shows what the tape is positioned for, where the puts and the calls are stacked, and which strikes saw a position appear today that did not exist yesterday, against a snapshot the tab saves for itself every day, so I run nothing, I click the tab and it has already fetched the chains and worked out what changed since yesterday. It describes positioning and it makes no call, and I do not make one from it either.

Short reads FINRA, the regulator that publishes the US short-selling data, and I keep its two datasets apart on purpose, because people mix them up. On the left is short interest, the actual open short position, settled twice a month. On the right is the daily short volume ratio, the share of each day’s volume marked short, where 40 to 50 percent is normal market-making and the only signal is the change against a name’s own average.

Capitol reads the trading disclosures that US senators and representatives have to file, filtered to the names on the book and the watchlist, with a lag column, because members get up to 45 days to report and the filing is old by the time you see it. Under it sit the members I track by name, and who to track is one small file.

Macro and Chain, the backdrop

Macro is 23 series from FRED, the St. Louis Fed’s free database, each one a card with a small two-year chart, and the economic calendar for the next ten days above them.

Chain is the supply chain map, and this one started as an edition. Bloomberg has a function that draws a company’s suppliers and customers, and in July I taught the agent to draw the same map from the filings alone, which is I Taught Claude Code to Draw a Bloomberg-Style Supply Chain Map from SEC Filings if you want the method, so on the desk it was one instruction: take the value chain I am interested in, lay it out layer by layer, what each layer buys and what it sells, put the listed companies under each layer, grade every link by whether the company disclosed it in a filing, said it on record, or it was only reported, and keep it priced. I am interested in the AI build-out, so mine are the AI chains for the two markets I trade, one screen each, and the frame and the hunt at the top are the research notes the map came out of. You can ask for any chain the same way, semiconductors end to end, backward and forward integration, whatever you are trying to understand, and the map comes out of the vault while the tab keeps it priced and in view.

So that is the floor, and every tab on it is there because I asked for it, which is what to take away, because you can build yours to your own requirement, whatever features you need, and it will look nothing like mine. What it looks like when you click into one company is next.

How the desk shows one company

Cmd+K from anywhere on the desk takes a symbol, and what opens is the page I used to assemble from five or six sites every time I looked at a name. This one is Alibaba, because it was on the watchlist with a busy insider month, and the company is only the example here, so read the page, not the stock.

The top: the market’s view

A chart, the valuation ratios, the quality numbers, and down the right what the street thinks, where the price sits in its 52-week range, and what the insiders have been filing. That right-hand column is the part I used to open three different sites for.

The middle: the filed numbers

The statements as filed, six years of them, income, balance sheet, cash flow, with the ratios, the segments, the estimates and the peers on their own tabs, and the earnings record underneath, every quarter against the estimate. I read the percent-change view first, because the shape of the change tells you where to look before any level does. This block is the one part of the desk that needs the paid feed, the statements, the peer set, the estimates and the seeds for the sandbox all come from it, and if you skip the feed this block stays empty and the agent pulls the same statements from the filings when you ask, the way the S&P Global run did, slower and one company at a time.

The bottom: a valuation sandbox

A discounted cash flow model, which is the standard way of turning a company’s future cash into a value today, with seven sliders, each one starting at the number the company actually filed, a fair value that moves as you drag, and a reset button that puts everything back. It answers one question, what am I assuming at this price, and never what anyone should do, and that is why it lives on the page and not in a spreadsheet, because the assumptions sit in front of you every time you look at a company and you cannot skip past them.

That is one company on the desk, and the page reads the same for any name you type.

How the desk measures risk

A screen full of live positions is not risk control, because concentration and leverage hide in plain sight on a green day, and I know that because the morning this tab first came up it told me two things I half knew and had never measured. So what I asked the agent for was the panel a risk desk at a fund puts in front of a manager, built for three accounts instead of a fund, reading all three books the way an outsider would, and this is what it built.

One headline card per book

Beta, volatility and the worst fall of the past year, each with a plain line under it saying what the number means on an ordinary day, then what a five percent index day would cost the book next to the free margin, then the cushion before a forced square-off, which is the point where the broker starts closing positions on its own because the margin has run out, and then where the book is concentrated by sector. Futures count at their full underlying value, so a levered book shows positions worth more than the account, and that is the point of showing it that way. The two accounts on the first desk run levered, about two times and three times, and the US book is unlevered and mostly cash, and the cards say so without me doing any arithmetic.

Do the books hedge each other?

A grid of how each book’s daily returns move against the others and against the two indices. Mine mostly do not hedge each other, the two accounts on the first desk move together at 0.87, and the US book is unrelated to both, which is diversification I got without planning it. That was the thing I half knew and had not measured, and now it is a number on a screen every morning.

What is inside each book

One line per position, sorted by size, with the share of the account, which can top a hundred percent with leverage, and the same beta, volatility and worst-fall numbers per name, and under each table a grid of which names move together. The first two lines are always the ones that decide the week, and that is the second thing this tab told me.

It is the six numbers an outsider would ask for first, on one screen, a long way from a fund’s risk system, and for three books that is enough.

What runs without me

The clock is the part of the desk I want to be careful about, because it is easy to picture more automation than there is, so here is exactly what runs on its own and what waits for me.

The desk’s own loops

The positions refresh every thirty seconds, the options tape on the index every ten minutes, and the heavy tabs, funds, filings, macro, every few hours to a day. None of that needs me, and none of it needs the agent either, it is the desk doing its rounds.

Nine alert rules

Every minute the desk checks nine rules against the data it already holds, and when one fires my Mac gets a notification and the desk gets a line at the bottom, once per rule per day. The rules are a day move on a name I hold or watch, one rule per list, margin used past a line, a futures contract inside a week of expiry, earnings inside three days, a congress filing on a held name, an insider cluster, and an activist filing on a name I own or watch. They are one small text file, and adding a rule is adding a line. They do not decide anything, they tell me where to look.

The rituals are skills

Then there are the jobs I fire with one phrase. “Run the watch” is the daily check on the model portfolio, prices, filings, news and the catalysts coming up, written as a dated note. “Run the tape” is the Flow tab as a written report, the same read on the names I own or cover, filed as a note I can come back to, and that is the skill from I Do Not Trade Options. So I Taught Claude Code to Read Them. if you want it in full. “Screen the inbox” sorts what I clipped that day. Each phrase is a skill, and the calendar tell and the selloff triage from the earlier editions sit in the same folder and run the same way. The desk shows, the skills read, and I fire them, because the two rules I run all of this on are that collecting and judging are separate steps, and that every step leaves something I can check that is not the agent’s word, a file on disk, a filing link, a row count.

The always-on option

All of this runs whenever my laptop is on, and for most people that is enough, the desk is up while you work and the alerts land on it. If you are serious about a desk that never sleeps, the whole folder, the vault, the agent and the desk, can live on a VPS, which is a small rented computer in the cloud that is always on. Obsidian syncs to it, the desk runs there around the clock, and Claude Code’s remote control lets you talk to the agent from your phone. I run it that way, and it changes two things, the alerts and the snapshots no longer depend on a lid being open. One more option, and it may matter to you more than to me: connect Telegram to Claude Code, and the ping lands on your phone the moment it fires, wherever you are.

What this desk is built for

A reader quicker than me is going to look at twelve tabs and ask what the catch is, so let me separate two things that look the same and are not, what this setup cannot do, and what I have set mine up not to do, because almost everything on the second list gets called a limitation and it is a choice.

It backtests off the broker’s history

Once the broker’s API is wired, the same connection that prices the desk hands you the price history as well, minute by minute, and that is enough for a backtest, which is running an idea over past prices to see whether it would have worked. I run one sometimes, the last one was an intraday options idea over a hundred sessions of the broker’s own candles, the agent built the test, scored it, and came back with the honest answer that the idea only works when one judgement call is right, which is exactly what a backtest is for, and a factor model or a full risk engine is one more description away if that is how you run money, because the history is already there.

It places orders through the broker

The same API takes orders on almost every broker, Interactive Brokers, Alpaca, Robinhood’s separate account for the agent, Zerodha and the other large brokers, so when a strategy tests well the desk can trade it, the agent sends the order through the broker, sized by the rules you wrote down, and the fill comes back onto the same screen. That is the part most people ask about first, and the answer is yes, if your broker opens that door, and it is the reason the backtest above is worth running at all.

The paid feed is a data provider

The financials block on the ticker page runs on the paid feed, and I want to be plain about what that is, because it is the same arrangement every screener you pay for has, the statements come from a data provider and the tool sells them back to you with a screen around them, and here the provider is the only thing you pay for and the screen is yours. And if you do not connect a broker at all, or connect it only for your positions, the quotes, the history and the charts come free from Yahoo Finance, delayed by about fifteen minutes on most exchanges, and it covers almost every market there is, which is exactly how the Global tab runs today.

What I do with orders

Now, what I actually do, because it is not one answer. For the intraday options trades, the straddles and the scalps, the broker’s own algo platform places the orders, which is the broker’s tool for firing orders by a rule you set, because that is a rules game where the entry and the exit are decided before the session opens and speed matters more than my hand, and I would be the slow part. For the equity book I do not use it, and I do not think I will, because that one moment at the broker, after the desk has put the whole picture in front of me, the size against my rules, the reference price, the cushion, is the part I like doing myself, and the decision has to be mine. So build yours whichever way you trade, and the desk does not care either way. Which agent sits at it does not matter either, and that is next.

Any agent can sit at this desk

You have the list already from the map at the top, Kimi Code, Grok Build, Codex, the desktop apps, Claudian if you never want to see a terminal, and any of them opens the same folder and gets the same twelve tabs. Claude Code is the main one here, on Claude Fable 5.1, and it runs inside Obsidian in a terminal tab like this.

Grok Build sits in the next tab, on Grok 4.6, and I use it most days alongside Claude for one specific reason, which is that Grok reads X live, and X is where most of the seasoned investors I follow actually post, so when I want to know what people are saying about a name that is the agent I ask, and the answer lands in the same vault as everything else. And 4.6 is a very good model on its own, so some days it is simply the one I open first. Two agents, one folder, one rulebook, and whichever company ships a better model next month gets the chair without anything else changing.

Kimi Code is the third one I run inside Obsidian, and I want to put one photo here rather than a screenshot, because it says more about the company behind it than any benchmark does. This is the office of Moonshot AI, the company that makes Kimi, and what is mind-boggling to me is that this is a team that closed a 2 billion dollar round in May 2026, led by Meituan’s venture arm, at a valuation above 20 billion dollars, and they are sitting in that room, on those chairs, building the model. That is an extremely, extremely good team, and if the Claude subscription is the thing that feels expensive to you, Kimi Code is the one I would try first, because it opens the same folder and reads the same rulebook, and your desk will not know the difference.

How to install the Claude Code build

This is the Claude Code build, the memory, the agent and the plugins, in the order you install them, and it is the part every tab above stands on. The live desk itself, the twelve tabs, is the section after this one, and it stands on everything here. If you want the install in full with every screen, How I Set Up Claude Code as My Investment Research Analyst still holds, with its walkthrough video, and so does the standing rule from that edition: if any step confuses you, copy the confusing part, paste it into the agent, and ask it to walk you through, because you are setting up an analyst and it can do its own tech support.

One: Obsidian and the vault

Download Obsidian from obsidian.md, open it, and create a vault, which is nothing more than choosing a folder on your computer and giving it a name. That folder is the whole build from here on.

Two: the agent

Claude Code installs in two or three steps, and the current ones live on its official quickstart page, so ask Google or any model, how do I install Claude Code, and follow that. Kimi Code, Grok Build and Codex install the same way, one search each. If you would rather never see a terminal, install the Claude desktop app or the Codex app instead and connect it to the vault folder, the notes in Obsidian and the agent in its own window.

Three: the plugins

In Obsidian go to Settings, Community plugins, Browse, and install three by name: Terminal, which opens the agent next to your notes, Local Images Plus, which keeps the charts inside anything you clip, and Claudian if you want the agent as a chat window instead of a terminal. Then Settings, Core plugins, and switch on Web Viewer, which is what puts the desk and any web page in an Obsidian tab. In your browser, add Web Clipper from obsidian.md/clipper, open its settings, add your vault, and in the default template set the note location to raw. Playwright, the browser the agent drives, is one sentence typed at the agent later: install Playwright for yourself.

Four: the folders and the rulebook

Open the Terminal tab, type claude, and paste this:

Create this folder structure for my knowledge base:

- raw/ - this is my inbox where I'll dump source material
- wiki/ - this is your domain, you'll write and maintain everything here
  - Create a _master-index.md inside wiki/ with the heading "Knowledge Base Index"
    and a note that says "Topics will be listed here as they are created."
- output/ - this is where query results and reports go

Just create the folders and that one file. Nothing else yet.

Then create a file at the top of the vault called CLAUDE.md and paste this in. It is the rulebook, the agent reads it at the start of every session, and it is the same one I published in the first edition, because it has not needed to change.

This file tells you how to maintain this knowledge base across every session.

## Knowledge Base Rules

- This is an LLM-maintained knowledge base. You are the librarian.
- The wiki/ folder is YOUR domain - you write and maintain everything in it.
  I rarely edit wiki files directly.
- raw/ is the inbox. When I dump files here, you process them into the wiki
  during a "compile" step.
- wiki/_master-index.md is the entry point. It lists every topic folder with
  a one-line description. Always keep this up to date.
- Each topic gets its own subfolder in wiki/ (e.g., wiki/ai-agents/) with its
  own _index.md that lists all articles in that topic with brief descriptions.
- Always use [[wiki links]] to connect related concepts across topics.
- When compiling raw material:
  1. Read the raw file
  2. Decide which topic it belongs to (or create a new one)
  3. Write a wiki article with key takeaways and relevant links
  4. Update that topic's _index.md
  5. Update wiki/_master-index.md
  6. If a raw file spans multiple topics, create articles in both and cross-link
- Keep articles concise - bullet points over paragraphs.
- Include a ## Key Takeaways section in every wiki article.
- output/ is for query results and generated reports.
- When answering questions, read _master-index.md first to navigate, then
  drill into the relevant topic _index.md, then read specific articles.
- When I ask you to "compile", process everything in raw/ that hasn't been
  compiled yet into the wiki.
- When I ask you to "audit" or "lint", review the wiki for inconsistencies,
  broken links, gaps, and suggest improvements.

If you run a second agent, add a file called agents.md next to it with one line in it, read CLAUDE.md, and every agent you open reads the same rules.

Five: clip, compile, ask

Open any filing or article in your browser, click the Web Clipper, and it lands in raw as a note with its charts. Here is the Web Clipper video, the clipping in motion, the same one from the Obsidian edition, because it is easier to watch once than to read twice.

Then type compile in the terminal and watch the wiki folder fill itself, one topic folder and one index at a time, and that is the memory working. From here the whole thing is three verbs. You clip. You compile. You ask, and when you ask, tell it to save the answer as a note, because every answer that gets saved is a note the next answer can read, and that is the part that compounds. For the first real run, take the prompt from the data section, the S&P Global one, change the company to any name on your own market, send it, and what comes back is your first sourced table.

Six: the first skill, and permissions

When a run comes back the way you want it, tell the agent to remember that job as a skill called analyse, and it writes the job down in a small file under the vault, and from the next time on you type the name and the company. On permissions, the agent asks before it touches a file or runs anything, and I would keep that on for the first week so you can watch what it does, then switch it off and let it work, because the vault is the only folder it can see and Obsidian’s File recovery plugin under Core plugins keeps a copy of every version of every note, which is the undo if anything ever goes wrong.

Three questions a subscriber sent

These came from a subscriber who has been running the setup since the earlier editions, so if you are new, skip this block for now and come back once your vault has something in it.

My research lives in OneNote, with a SharePoint folder per company. How do I move it into Obsidian? This is the hardest step in the whole piece, honestly, and I am not going to pretend otherwise. OneNote does not hand you a notebook as plain text, it exports a section at a time as a PDF or a Word file, so a few thousand notes is a slog of an afternoon, or two, and SharePoint is easier only because it is already files, you copy each company’s folder as it is. So what I would do is not move everything at once. Move the ten companies you are actually working on, into raw under each company’s name, type compile, and let the agent read the PDFs and the documents the way it reads a clipping and write each company into the wiki, and then move the next ten when you need them. And the reason to go through it at all is what you get at the end, which is that every one of those notes becomes plain text the agent can read, so when you ask it about a company it has years of your own thinking on that company in front of it, all of it, and no tool you pay for can give you that.

Should each stock get its own raw, wiki and output, or one of each? One of each, and the company lives one level down. raw is an inbox, it is meant to be emptied, so nothing in it needs a company folder for long. The wiki gets one folder per company, and that is where the stock-level view lives, the filing notes, the concall notes, the thesis. output gets one folder per company too, with a quarter folder inside it for the reports. That is exactly how mine runs, and three inboxes per stock would give the agent thirty maps to read instead of one.

The graph view is full of _index and company_overview dots. How do I make it show relations? The graph draws one dot per note and one line per link, so it can only show the links that exist, and right now the index notes link to everything, which is why they crowd the picture. Two fixes. In the search box on the graph, type a minus sign and the name of the note you want hidden, so -file:_index for the index notes, and they are gone. Then tell the rulebook what a link is for: add one line saying every company note links to its sector, its peers and the filings it came from, and from the next compile the graph draws the relations you care about, because the agent only draws the links you told it to.

That is the Claude Code build, and every reader of the earlier editions already has it. The desk on top of it is next.

How the desk was actually built

The honest answer to how this got built is that most of it was built once already, in a different place, and I should tell you that story first, because otherwise the speed of it makes no sense. Last year I built GreekSoup, which was a research dashboard of my own, and at the time I was planning to launch it commercially later on, because that was the mood everyone was in, right, AI is going to kill the whole software industry and anyone with a framework and a weekend can ship a product, and I had a framework. A few things were on my mind, though. I am an investor first and a builder of experiments second, and once you launch an app you are doing the marketing and the product and the support and the pricing, and all of that would have turned my life into something I did not want it to be, and that was one of the reasons I dropped the idea, kept the software for my own work, and moved on. I did not write the code of it myself, but every feature in it came out of my own research framework, the way I read a company, the ratios I look at first, the peer table, the valuation sandbox, so many of those things were mine already, and when this desk started, what I actually did was go back into GreekSoup, take a screenshot of a feature I had already designed once, hand it to the agent and say I want this here, and it came back in minutes, because a picture of a thing you have already thought through is the best spec there is. That is where the charts, the statements, the ratio groups, the peers and the DCF came from, and the whole desk is yours to download at the end of this piece.

The desk itself started from one page in the vault. The note said what to build against the broker’s API, and it had one rule in it, prove the connection on the cheapest piece first, the holdings, and the agent read that note and had both accounts on one page before I went to bed, which I did not expect. From there it was a loop that never really stopped, where I sent a screenshot in, a screen came back, I said what was wrong with it, and it went again.

If you are starting from nothing

If you do not have software of your own to carry across, and most people will not, start with the apps and the websites you already like. Say you use Koyfin, or Finviz, or Trading Economics, or your broker’s own terminal, and on each one there is a feature you would miss. Go there, take a screenshot of it, paste it into the agent and say I need something similar here, and it comes back in a few minutes, and then you do the next one. Five or six of those on one screen and you will start having opinions about what is missing, and that is when it becomes yours. And the simplest version of all is this piece. Clip the whole article into your vault with the Web Clipper, open the agent on it and say build me this for my broker, my market and my names, and then screenshot the things you want changed.

Give it the keys

The second thing is the keys, and this one is simple. You give it your broker’s API key, and the feed’s key if you take one, and before it builds anything you tell it to find out what each key can do, because you do not know, and it can check in a minute. Mine went through everything the feed can be asked for on its own and came back with a note on which of it my plan allowed and which was gated, and wherever the feed was gated it went to the free record instead, EDGAR for the fund filings and the activist filings, CBOE for the options chains, FINRA for the short data, and the broker’s own history for everything on the first desk. It can reach a lot more than you think, and you only find out what it can reach by using it, which is why I would use it for your research for a few months first and ask for the desk after, because by then you know what to ask for.

Ask for the premium version

The third thing is the look, and it matters more than it sounds, because a screen you enjoy looking at is a screen you actually use. Tell it to act as the designer of a premium terminal, or as a visualisation specialist, the kind a fund pays a lot of money for, and to build to that standard, a premium dark dashboard if that is your taste, and mine came back with the sidebar, two themes, a dark one and the light one you see here, and the search box that jumps to any ticker from one ask like that. From there it is screenshots the other way, every screen sent back with what was wrong on it, and some of it came back wrong, because for a day the desk read one account’s margin as far safer than it was, it was reading the wrong line in the broker’s reply, and I only caught it because the broker’s own page sat next to it and the two did not match. A screen that is confidently wrong is worse than no screen, so the desk still sits next to the broker’s page, and I still look at both.

What runs it, in one paragraph

One small program runs the whole thing. The agent wrote it in Python, which is the language it is most fluent in, there is nothing else to install and nothing to host, so there is nothing to pay for beyond the feed if you take one. The lists you saw, the watchlists, the funds, the members, the chain, are small text files you can open and edit. Everything it fetches once, it keeps, so the heavy tabs open instantly the second time. The agent wrote every line of it, and I have not written one.

Your first tab

So if you want yours, you will not start with twelve tabs. You will start with the one screen you look at first thing every morning, you will screenshot the thing you want it to look like and hand it over, it will come back wrong in two or three places, you will say what is wrong, and it will fix it, and the second tab is one more screenshot, and by the end of the week it adds up to the desk you had in mind. What is left is what all of this costs.

What the whole desk costs

Go back to the tab stack at the top of this piece for a second and add up what yours costs a month, because the number is bigger than you think. This desk runs on Obsidian, which is free, five plugins, which are free, one agent subscription, which you are paying for already if you use any of this, the broker’s API, which most of the brokers I named give their own clients for nothing, and the public record, which has always been free. The one thing on the list with a price is the feed, and it is optional, about 29 dollars a month, or 19 if you pay for the year. So the honest cost of a small Bloomberg for one desk is one feed you can skip, and the time it takes to describe what you want.

The whole desk, without my positions, is at the link below, on GitHub, which is the public site where code lives and anyone can download it, one account wired, the broker connection written for mine and adaptable to yours, and if you like it the way it is you can start from there and change what you want.

Opensource Github Repo

If you want to see how the older versions of this setup looked: How I Set Up Claude Code as My Investment Research Analyst was the first install, How I Set Up Claude Code as My Investment Research Analyst 2.0 was the rebuild inside Obsidian, and Claude Code for Investment Research (Part 4): The Obsidian Layer That Makes Everything Compound is the clipping and the memory in full.

The video walkthrough goes out over the next couple of weeks, and if you want it, and the next build after it, make sure you are subscribed. And if you build yours, or take this one somewhere I have not thought of, tell me in the comments, because the best setups I have seen came from readers who did exactly that with the last edition.

The whole desk turned out to be a folder on my Mac and a few sentences about what I wanted to see, and the sentences are the only part of it you cannot download from the link above.

*Disclaimer: Alpha with AI covers US-listed and other global securities for educational and research purposes only. Nothing here is investment, legal, or tax advice. The workflow and its outputs read public filings, investor calendars, and news sources; outputs can contain errors and are observations of past filings, not predictions or recommendations. AI tools, including Claude Code, were used in the research and drafting of this edition. Always verify against primary sources and consult a qualified professional before making investment decisions.*

Written by Shubham Borkar

Financial Clarity. Insightful Ideas.

Delete your old finance skills.

Not the one you built last week. The old ones. The skill you wrote in March, patched in April, then tightened again in June when a new model came out and the output changed under you. The one you have edited so many times you stopped counting.

I can hear the objection already, because it was mine. Why would I delete a skill I have been updating with every new model anyway? That file has survived three model generations. Every fix in it was paid for by a bad output on a real company. Delete it and start fresh, and some nuance in there is going to get missed.

Hold that objection. It is correct. It is also the reason most people never see the real problem.

Open one of your old skills right now and read it top to bottom. Count the rules you cannot explain. You know you wrote them. You cannot say what they were for. Right?

Every one of those rules went in on a specific day, to stop the model you had back then from doing something that annoyed you. Maybe it hedged when you wanted a verdict. Or it kept estimating numbers the filing never disclosed, so you wrote a line that said never estimate, only quote. The line worked. You moved on.

Then a better model came out and stopped doing the thing.

The line stayed.

Nothing removes it, because nothing ever does. A rule that quietly died and a rule that is holding your entire output together look identical from the outside. The only way to tell them apart is to pull one out and run the skill again, and nobody does that. I did not either.

So the file only grows. Ten edits, twenty edits, three or four models, six months of honest maintenance, and what you own at the end of it is a working skill buried under fixes for problems that no longer exist. A patch log.

Before you decide I am telling you to select a year of work and press delete, I am not. There is one step that makes the whole thing safe, and I learned why it matters the hard way, because before writing a word of this edition I sat down and read my own oldest skills, top to bottom, rule by rule, and what I found in them changed what this edition is about.

We will get there. First, the reason updating never fixes this on its own.

Why updating an AI skill is not the same as rebuilding it

Think about how a rule actually gets into a skill file. Nothing about it is planned. The model gives you a bad output on a real company, on a real morning, and it costs you twenty minutes. You trace what went wrong, you write a line to stop it, you run it again, and the output comes back clean. The rule has earned its keep.

Now think about how a rule leaves the file.

It does not. There is no moment for it. A bad output triggers an edit, but nothing triggers a deletion, because a stale rule does not announce itself. It does not error. Nothing in the output points back at it. It just sits there, either doing nothing or quietly making the output worse, and from where you sit both look exactly like a rule that is working.

So the file moves in one direction. Every model release adds a few fixes and removes none, and the skill slowly fills up with corrections for the habits of models that are no longer running.

Fine, you say. Then I will have the model clean its own house. Hand it the file, tell it a new model just shipped, ask it to improve the skill.

I have done this many times. Here is what actually happens. The model works on what is in front of it. It tightens your wording, reorders your sections, patches the gap you pointed at, and hands back a file that is genuinely better than the one you gave it. What it will not do, not once in my experience, is tell you that half the structure exists to manage a model from March and the March model is gone. You did not ask that question, and the file sitting on the page keeps it from ever coming up.

That is the difference, and it took me longer to see than I want to admit. An update can only change so much. It polishes the answer to a question I asked months ago. Starting fresh is the only thing that goes back to the question itself.

I am not guessing at any of this.

By the time these skills reached everyone, I had already built mine, broken them, and rebuilt them for months. I walked through the full discipline of building one properly, including the four questions I answer before writing a single line of a skill file, in Learn to build a Claude skill for equity research. Start there if you have never built one, because everything in this edition assumes a skill worth saving.

The skills from those beta months have been through more updates than I can count.

Which brings me to this month, and to the person who finally pushed me to look at them properly.

What the creator of Claude Code says to do with your old skills

For the last couple of months, some of my oldest working skills had been coming back with weaker output than they used to. Not broken. Just not as good as a month ago, two months ago.

I blamed the model. Every time results come back worse than they did two weeks back, people blame the model. I was one of them. I even said it out loud on social media.

Then this month I ran into the creator of Claude Code saying the opposite. Boris Cherny, in a July interview at Y Combinator:

Every six months, delete your CLAUDE.md, delete your skills, delete your hooks. See what the model does.

And the part that pushed me harder than the advice. For Opus 5, his team deleted 80 percent of Claude Code’s own system prompt. His reason, in his words: a lot of it “was correcting for these behaviors that the model should have known, but it didn’t. Now Opus 5 just does it.” They do this on every release. A new model comes out, and they delete and rewrite a chunk of what they had.

That stopped me. The team closest to the model does not trust its own year-old instructions. I was trusting mine completely, and blaming the model for the results.

I will be honest here. I had a little bit of this idea myself, and I was not pushing it, because everything was working. Sometimes you need someone who deleted 80 percent of their own work to push you. I needed that push.

So I did what I asked you to do at the start of this piece. I opened my everyday skills, the ones that run my actual research, and read them top to bottom, rule by rule.

Here is the thing about any skill you actually use. The history is always there, sitting inside the file. No skill gives you the output you want on day one. It makes mistakes, you correct it, it makes new mistakes, you correct those, and update after update it adds up to the output you had in mind. That is true for every skill I run daily, and it will be true for every skill you build after reading this.

So the history was never the problem. I read those files and the history is all there, every fix I ever made. What is not there is a single reason. The file remembers what I changed. It does not remember why. Every rule I could not explain is a decision I made once, against a model that may not even be running anymore.

The reasons survived somewhere else, though. Somewhere I was not looking. That is where the rebuild starts.

A small note, and the only sponsor here is me. I keep getting messages asking how I set up Claude Code for stock research, beyond what the editions can teach. So I am opening a small consulting program: I set your system up with you, then support sessions over the following months as your real problems show up. I am doing a lot of other things, so seats are very few. If that is you, the waitlist is on the website or reply to this email.

Before you rebuild an old skill, find where its reasons were written down

Every rule in every skill was explained once. Not in the file. To a person.

Think about how a skill gets into anyone’s hands. Nobody installs a file of instructions on blind trust. Somebody had to convince you first. If you downloaded a skill from one of my editions, you read the piece before you installed it, and that piece walked you through why every rule was in there. If you built your own, you did the convincing somewhere else. The message where you explained it to a colleague. The note where you talked yourself into building it.

The reasons never make it into the file, because the file does not need convincing. You do. So the reasons sit wherever the convincing happened, and they are still sitting there today.

Let me show you what I mean with one of mine. And let me explain the skill properly first, like you have never read me before, because the example is useless if the skill does not make sense.

After a company announces its quarterly results, the numbers are public. Everyone has seen them. What nobody can see in the numbers is whether the people running the company believe in what comes next. So I watch what management does in the six weeks after the announcement. A team that just delivered a great quarter books itself everywhere. Conferences, investor events, analyst calls. They want to be in the room, because every question is now a question they can answer. A team that just delivered a bad quarter books nothing. Not because the numbers are secret. Because every room is now full of questions they do not want to answer. I am not predicting the next quarter with this. I am reading whether the insiders are behaving like things are fine or behaving like things are not, and that behaviour sits on public pages weeks before any analyst changes a number. I call this skill the calendar tell, and I taught the whole build in I Taught AI to Read a Company’s Earnings Before Wall Street Does. The Tell Is on Its Calendar.

Inside that skill sits one rule. Do not judge this quarter’s calendar on its own. Compare it against the same six weeks after each of the company’s last four or five results. If you opened the skill file today, that is all you would find. One line. Follow it or do not.

But when I published the skill, I had to convince you that line was right. So the full reasoning is on that page: some companies are simply busier than others, some have a heavy season and a quiet one, so counting events tells you nothing until you compare the company against its own normal.

That highlighted paragraph is what the file never held. The file has the rule. The page where the skill was first explained has the reason.

So here is the move, and it works whether you built your skill or downloaded it. Before you rebuild anything, go back to where the skill was first explained to a human being. For a skill you took from these pages, that is the edition it came from. For a skill you built yourself, it is your own note, your own message, wherever you once did the convincing. Collect that explanation. It is half of the skill, and it is the half the file cannot give you.

But an explanation is not a skill file. I cannot install an article. To turn the explanation back into a working skill, you need to know what you are looking at inside the old file, because not everything in there deserves to survive. Four kinds of things live in every skill file, and they age at completely different speeds.

That sort comes next.

Read more

Alpha with AI is now my full time work.

I have stepped back from client projects. A few existing engagements continue, and
everything else has stopped. From this month my entire working week goes into this
publication: reading filings, tracking the macro picture, and researching businesses
one at a time.

I have never laid out in one place what the paid side of this publication actually is.
That is what this post does. There are also two things landing in the next two weeks
that are worth knowing about before they arrive.

HOW I INVEST

Most analysts I have worked alongside took the standard route: theory first, the
certifications, then markets. I came the other way around. I was investing my own money(parent’s money if I’m honest) before I ever studied the discipline formally, and I later managed client money in direct equities. That order shaped how I work. I care less about elegant models and more about whether I can defend a position when it trades well below my cost.

Three questions decide every piece of research I publish.

Is the macro with the business or against it? Interest rates, the capex cycle, regulation, trade policy. A good business in a hostile cycle can stay cheap for years,
and I want to know which side of the cycle I am standing on before I read anything else.

Can earnings actually grow from here? Not the story, the arithmetic: the revenue
drivers, the margin structure, what management guided and what it then delivered. Over any period that matters, share prices follow earnings.

What is the business worth against what the market is asking? I value the whole
company, set it against the price, and insist on a margin of safety wide enough that I
can be wrong about something and still come out whole.

And one filter on top of all three: I only hold businesses I am comfortable holding
through an earthquake. If a falling price would shake me out of a position, the
research behind it was never finished.

WHAT PAID SUBSCRIBERS GET

3-4 AI workflow editions a month. The research systems I use to
produce all of the above. These deserve their own section, just below.

2-3 Research Dives a month. One company per dive, US-listed and other global names.
I read the filings myself, rebuild the numbers, and put a value on the business with
the working shown. Each dive ends with a decision: I am buying it in the model
portfolio, I am waiting at a lower price, or I am declining it. The declines get
published too. Knowing why I walked away from a popular name is worth as much as
knowing what I bought.

The equity model portfolio. Every dive decision is executed in the model portfolio
on publish day and stays on the record, wins and losses alike. You can audit any
position back to the research that created it. This is the book today:

Every buy and sell sits in one running ledger, oldest first, linked to the dive that
caused it:

The watchlist. For every company I have fully researched, my valuation and the
level where I would add are on record before the market moves. When a selloff comes,
the decision is already made. Below the researched names sits the radar: the companies
waiting their turn for the full work:

Model Portfolio

Subscribe now

WHAT AI PRIME ACTUALLY IS

The workflow editions need their own explanation, because the word “prompts” undersells them badly.

I am a fundamental investor who has used AI in my research every working day for more than two years shamelessly, and the workflows are where the two disciplines merge: investing principles first, AI built around them. Each published edition is the survivor of a longer process. I take one research problem, experiment with several approaches to it, burn millions of tokens testing them against real filings, and throw away the versions that break. What gets published is the one that held up, end to end, with the exact prompts.

So what you receive is not a prompt collection. It is a tested research system with the
failed experiments already paid for. I have read what gets published on this platform
about AI and investing, and no one is doing this work the way it is done here. That is
also why this tier is priced the way it is: 3-4 new editions arrive every month, and
the library already holds more than forty.

Workflows like this: take your pick.

SIX COMPANIES WENT ON THE WATCHLIST THIS WEEK

A company earns a watchlist row when I can name the specific thing the market is
worried about and the evidence that would settle it. Six businesses earned one this
week. The names are in this week’s paid note. The businesses, so you can see the ground this list covers:

• an ice cream company that owns about three million freezer cabinets standing in shops it does not own

• a company that gets paid for decades every time certain jet engines come in for
  servicing

• the company that built the weight loss drug market, now playing catch-up in it

• a Japanese maker of the miniature ball bearings that robot hands will need

• the maker of the unglamorous chips that run the locks, the battery and the radar in your car

• an armoured truck company that collects cash from shops and ATMs

The paid note covers all six: what each business does, what broke, and the number I am waiting on before the full research begins..

THIS WEEKEND

A new AI post is on the way. It is an important one, especially if you have implemented at least one of my workflows. More on that when it ships.

THE BOOK IS JUST GETTING STARTED

The model portfolio is six weeks old. Four positions, and more than 90% still in cash.
That cash is about to go to work. The dives now come at full-time pace, and over the
coming months I intend to deploy most of that cash, possibly all of it, into businesses
that pass the three questions above. Steadily, and aggressively when the research earns
it.

Whatever the pace, the care stays the same. We are in a bull market, and I treat it as
a prudent bull market. There is an old rule about being careful about who you go into business with. Buying a share is going into business with a company, and every name that enters this book passes through exactly that level of care.

Joining now means being in at the start of the record, with the cash still waiting and
every decision published with its reasoning as it happens.

One more thing, briefly. Most of my own investing over the years has happened outside the US, and only a very small portion of my personal portfolio was in the US market, so I largely missed the US rally of the last two years. Apart from my other portfolios, I have now started managing a new family account in the US market, where I will deploy slowly and steadily. Paid subscribers will be able to see those positions too, kept fully separate from the model portfolio.

WHAT IT COSTS

This post is free in full, so you can judge the work before paying for any of it.

A word on these numbers. I restructured the offering this month. The research tier,
the dives, the model portfolio and the watchlist, is now priced at a fraction of what
this work cost here earlier. That drop is deliberate but it is not permanent.

AI Prime is the exception, for the reasons laid out above. Its price does not drop.
What it carries today is a temporary reduction, and it returns to its original price
soon. If the workflows are what you came for, the arithmetic favours joining before that happens.

Here is what the next two weeks hold on the paid side: the full note on the six
companies above, the model portfolio acting on its written levels if the market touches
them, and the AI post. The free half of every edition will keep reaching you either
way. The story is in the free half. The numbers, the valuation and the decision sit in
the paid half.

If you have been reading the free halves for a while, you already know whether this
work is for you.

Subscribe now

*Disclaimer: Alpha with AI covers US-listed and other global securities for
educational and research purposes only. Everything in this edition is my personal view
and, where stated, my own portfolio action. None of it is investment, legal, or tax
advice, a recommendation to buy, sell, or hold any security, or tailored to anyone’s
situation, and I am not a registered investment adviser or research analyst. The
valuations here are my opinions with the method shown, not price targets or calls to
act. The model portfolio is a model book, tracked on notional capital, and is not a
record of real trades. I may hold or take positions in the securities discussed, my
views can change without notice, and when they do, you will read about it here. Every
figure traces to a primary public filing, accession numbers on file, but filings can be
misread and models can be wrong, so verify against the source before acting on
anything. AI tools, including Claude Code, were used in the research and drafting of
this edition, with every number checked back against the filed documents. Investing
carries risk, including the loss of capital. Do your own research and consult a
qualified professional in your jurisdiction before making investment decisions.*

Financial Clarity. Insightful Ideas.

I have learned to distrust a stock that looks perfect.

That distrust is earned. I have paid for the lesson more than once, buying the clean-looking thing and finding out later what the clean was hiding. So now when a business screens as the best in its industry and trades like one of the worst, I do not get excited. I get suspicious. A gap that wide is not a gift someone left out for me. Somebody on the other side of it knows something, and the work is finding out what.

This one screens close to perfect.

Margins at the top of its entire global industry. No real debt. A cash balance bigger than the business needs. Almost all of its profit handed back to shareholders every year. Number one in every market it operates in, not most of them, all of them. And it trades for a fraction of what the big Western name in the same business commands.

Read that list and you want to own it. I wanted to.

Then you find the part the list leaves out.

It does not own its own technology. The system that moves the drivers, the payments, the data underneath all of it, none of that belongs to the company. It rents it. From the parent. The same parent that holds most of the stock and nearly every line of the software, and that is, right now, being bought by the company’s single biggest competitor in the world.

So the clean screen is not clean. The business is genuinely very good, and that part is not really in question. What is in question is whether a small shareholder standing on the outside ever gets to collect on it, or whether the people holding the controls quietly decide that for you.

The company is Talabat.

The odds are, probably you have never heard of it unless you have spent some time in the Gulf, in which case you have probably used it. It is an app which large part of the Middle East opens when it wants food, groceries, or medicine at their doorstep. It has traded in Dubai since December of 2024, which is carved out of a German parent that has itself become a takeover target since.

I wanted this to be a simple one. It is not. And working out why is the whole point of what follows, because the thing that decides your return here may turn out to have almost nothing to do with how good the business is.

What Is Talabat, and Why Is the Best Business in Food Delivery Trading at a Discount?

(The whole thesis up front, for anyone who reads nothing else.)

What it is. Talabat (DFM: TALABAT) is the largest food-delivery and quick-commerce company in the Middle East and North Africa, the app people in eight countries open to get restaurant meals, groceries, and pharmacy orders to the door. It listed in Dubai in December 2024. Delivery Hero, the German group that built it up, still owns about 80 percent.

The one fact that matters. Out of every dollar that flows across its platform, Talabat keeps about six and a half cents as operating profit, on the adjusted measure the company itself forecasts and reports. That is among the highest margins of any listed food-delivery company, and more than double DoorDash’s. Even Eternal, the Indian company behind Zomato and Blinkit that Talabat is most often compared to, earns a little less on its food-delivery business than Talabat does. This high-margin food business is the engine, and it throws off cash: no debt, around 800 million dollars of net cash, and a dividend that returns about 90 percent of each year’s profit to shareholders. That last figure is a payout ratio, not a yield. The company hands out most of what it earns, which at today’s share price works out to roughly a 5 percent dividend.

Why it looks cheap. The market values the business at about 12 times a year of that operating profit. DoorDash, its closest Western peer, trades at 18 to 24 times while earning less than half the margin. India goes further still: in their sum-of-the-parts valuations, analysts there put Zomato’s food-delivery arm alone at something like 35 to 45 times a year of its profit, part of the steep premium Indian markets pay new-age technology companies. That Indian number is not a fair yardstick for Talabat, and I am not holding it up as one, but it shows how differently the same business gets priced depending on where it trades. So the best-margin, cash-richest name in the industry is also one of the cheapest. On the numbers alone, it looks like a high-quality business on sale.

The catch, and it is a big one. Talabat does not own the software that runs it. The system that dispatches the drivers, the payments, almost all of it is licensed from its parent, Delivery Hero. And Delivery Hero is itself being bought by Uber, in a deal expected to close in the second half of 2027. So control does not simply pass to a new owner. It passes to Talabat’s biggest global competitor, the company behind Uber Eats. Talabat will still not own its technology afterwards. The licensor just sits inside Uber instead of inside a German group, and whether that licence continues on the same terms has not been spelled out anywhere. There is a version of this that cuts the other way: once the company that owns Talabat’s software sits inside the same group that controls Talabat, that owner has its own reason to keep the technology healthy, because it now profits when Talabat does. Whether that comfort or the competitive conflict wins out is genuinely open. For now Talabat keeps its brand, its Dubai listing, and its minority shareholders, who are not being bought out. What Uber does with it later, leave it standalone, merge it with its Careem business in the region, or fold it deeper into its own system, is an open question. And because of the way Talabat is legally set up, a small outside shareholder has no guaranteed fair-price exit if the controller ever decides to buy them out. The discount is not the market being slow. It is the market pricing all of that uncertainty.

What could turn it. Talabat stays cleanly listed under Uber with the dividend intact; margins settle as the current spending phase ends; the subscription and grocery arms keep scaling; or a buyer takes the whole thing private at a real premium.

What would break it: Keeta, the Chinese-backed challenger, winning the Gulf price war; a new owner cutting the dividend that is much of the reason to hold the stock; or a cheap, forced buy-out of the minority.

How Did Talabat Go From a Kuwaiti Startup to a Delivery Hero Company Trading in Dubai?

Talabat did not begin as anyone’s crown jewel. It began in 2004, in Kuwait, as one of the first websites in the region that let you order food without picking up the phone. The name is the Arabic word for orders. At the time fewer than a third of Kuwaitis were online, so this was an early, almost premature bet on people doing ordinary things through a screen.

Who exactly founded it is not settled, and I am not going to pretend it is. A group of Kuwaiti founders got it going; the name most tied to its early rise is Abdulaziz Al Loughani. What matters more for this story is what came after, because Talabat spent the next decade being handed up a chain of larger and larger owners, and control moved a little further from the people who actually used it at every step.

It changed hands at least once while it stayed Kuwaiti. Then, in February 2015, it left the region entirely. Rocket Internet, the Berlin startup factory, bought Talabat for about 170 million dollars. At the time that was the largest technology exit the Middle East had seen since Yahoo bought Maktoob. Weeks later Rocket folded Talabat into Delivery Hero, the German delivery group it was backing, and Delivery Hero has owned it ever since. What Delivery Hero itself paid was never disclosed, because it was not really a purchase. It was one Rocket company being moved inside another.

From there, Delivery Hero did something deliberate.

It made Talabat the master brand for the whole region and fed its other Middle East businesses into it. It bought Carriage, a Kuwaiti rival, in 2017 for around 100 million dollars, ran it on Talabat’s system, and let the Carriage name fade. It folded in Otlob, its Egyptian brand. In 2019 it bought Zomato’s business in the United Arab Emirates for 172 million dollars and put that under Talabat too. In 2020 it paid 360 million dollars for InstaShop, a Dubai grocery marketplace, though that one it kept running on its own. One market sits outside the umbrella: Saudi Arabia, the largest economy in the Gulf. Talabat was there once. It sold food in Saudi Arabia in its early years, and Delivery Hero’s 2015 takeover announcement still counted Saudi Arabia among its strongest markets. But Delivery Hero also owned a Saudi delivery brand of its own, HungerStation, and in the consolidation that followed, Saudi Arabia became HungerStation’s territory. Today Talabat does not operate there at all, and its executives describe staying out as a choice about where the money earns more, not a door that is closed. Worth remembering the next time you hear it called the leader of the region.

While the ownership was being reshuffled at the top, the business itself grew into something wider than food. Talabat spread across the Gulf from 2011 and into Egypt and Jordan soon after, settling at the eight countries it covers today. During the pandemic in 2020 it launched talabat mart, its own grocery service that promises delivery in about half an hour from small local warehouses the industry calls dark stores (an unglamorous name for a shop with no shopfront, built only to be picked from). From 2023 it began selling a paid membership, talabat pro, that bundles free delivery and discounts. The price is AED 29 a month including tax, about 8 dollars, and in Dubai a single delivered meal often costs more than that. So the fee is not the business. The behaviour is. A customer who pays even a small monthly charge for free delivery has less reason to open four rival apps to compare, because an order placed anywhere else forgoes something already paid for. The membership turns an open contest into a default, and it hands Talabat a piece of revenue that repeats every month whether or not you order that day. The company discloses enough to show how far this has gone. At the IPO, about 8 percent of its customers held a subscription. By the middle of 2026, subscribers accounted for 51 percent of everything ordered on the platform by value, up from 37 percent a year earlier. The whole strategy has a name inside the company: the everyday app. The idea is to stop being the thing you use when you want a takeaway and become the thing you open for groceries, pharmacy, and the rest, several times a week instead of once.

For most of the run-up to the listing, Talabat was run by Tomaso Rodriguez, chief executive from 2019 to 2025, who oversaw the consolidation and took the company public. In October 2025 the company said he would hand over to Toon Gyssels, who took the job on the 21st of November 2025. Gyssels is a returnee, back after earlier stints at Talabat, and he has framed his tenure around leaning harder on AI. A change at the top is not, on its own, a warning sign. It is worth noting all the same that the person who built the story now sits on the board, and someone new owns the results.

Which brings us to the part that actually put Talabat in front of you. On the 10th of December 2024, Delivery Hero listed it on the Dubai Financial Market under the ticker TALABAT. By the money it raised, it was the largest technology IPO in the world that year, though 2024 was a thin year for big technology listings, so the label flatters it a little. The offer price was 1.60 dirhams a share, which valued the whole company at roughly 10.1 billion dollars. Demand was strong enough that Delivery Hero raised the size of the sale from 15 percent of the company to around 20 percent, and took in about 2 billion dollars.

Here is the detail that tells you whose deal this was. Every one of those dollars went to Delivery Hero, not to Talabat. It was what the market calls a secondary offering (the existing owner selling its own shares, rather than the company issuing new ones to raise money for itself). Talabat received nothing beyond its expenses. Delivery Hero cashed in part of its most valuable asset, kept about 80 percent, and kept control. And there is one more piece of the structure that matters a great deal later, so it is worth slowing down for. The shares you buy trade on the Dubai Financial Market, the stock exchange in Dubai. The company itself, though, is not registered in Dubai. It is registered in the Abu Dhabi Global Market, which is not a stock exchange at all. It is a financial free zone in a different emirate, Abu Dhabi, that runs on English-style common law, with its own courts and its own company law. So you trade the stock under Dubai’s exchange, while your rights as a part-owner, how you vote and whether you can be forced to sell your shares, are set by Abu Dhabi’s rules. Two emirates, two different jobs. Hold that. It becomes the whole question near the end.

The listing itself was not a triumph on day one. The stock opened above the offer price, then reversed, and closed its first session at 1.49 dirhams, nearly 7 percent below where it had been sold. A great business, a record IPO, and a market that was not quite sure what to make of it from the very first afternoon.

How Big Is Gulf Food Delivery, and Why Is It the Most Profitable in the World?

That is the company’s story. The market it lives in has one too, and it starts with a number that does not exist.

No one publishes the size of the UAE food-delivery market. Not the government, not the exchange, not Talabat. The private research firms that estimate it do not agree, and their numbers run from about 700 million dollars to more than 9 billion for the same year. That range is not a real disagreement. It is that they count different things. The low numbers count only the fees the platforms keep. The high numbers count the full price of every order, the food and the fees together.

Since there is no official figure, I estimated it from Talabat’s own numbers, and the steps are simple enough to check. Talabat processed about 9.4 billion dollars of orders across its eight countries in 2025. Assume the UAE, its largest single market, is somewhere between 28 and 35 percent of that, which is what the disclosures hint at without stating: call it 2.6 to 3.3 billion dollars of UAE orders. Outside estimates put Talabat’s share of the UAE market near 42 percent, so divide, and the market comes out at roughly 6 to 8 billion dollars a year counting the full value of orders, stretching toward 9 if either assumption is loosened, or 0.7 to 2.5 billion if you count only the platforms’ fees. It is growing at a high single-digit to low double-digit rate. Three things to be honest about with that estimate. The 42 percent share comes from an unaudited restaurant-industry panel. The UAE slice of Talabat’s own business is my assumption, not a company disclosure. And the 9.4 billion I started from is everything ordered through Talabat, groceries included, while the 42 percent share was measured on restaurant orders alone, so the honest label for my range is delivered food and groceries, not food by itself. Treat it as a careful estimate, not a hard fact.

Here is the part that matters more than the size. Judged by what its biggest player earns, I cannot find a more profitable food-delivery market anywhere.

Talabat keeps about six and a half cents of profit out of every dollar ordered. DoorDash, in the United States, keeps under three. The reason is a gap the Gulf has and almost nowhere else does. On one side, the baskets are large: an affluent, convenience-first population orders expensive food. On the other, the cost of carrying it is low, because a deep pool of migrant labour drives the bikes. High value per order, low cost per order, and the space between the two is the margin. Then stack on the physical setup. This is the densest high-rise living in the region, so a rider drops five orders in one tower instead of driving between five houses. And the heat runs extreme for close to five months a year; for three of them, mid June to mid September, the government bans outdoor work in the early afternoon. All of it pushes people indoors and onto the app. Wealth, density, cheap delivery labour, and heat that keeps people indoors and ordering. That is why an eleven-million-person country supports a multi-billion-dollar delivery market and the highest margins in the business.

It is worth being just as clear about the other side of that, because it sets the ceiling on everything that follows. The UAE is small. India, where Zomato and its rivals operate, has about 128 times as many people. The five Gulf countries Talabat serves hold about 26 million people between them, fewer than greater Delhi. So Talabat cannot grow the way an Indian platform grows, by adding a hundred million users a year. It is already the largest player in every market it operates in. Its growth has to come from somewhere narrower: defending and taking share from new entrants, selling more than restaurant meals through the same app, getting the customers it already has to order more often, and leaning on its faster-growing smaller markets in Egypt, Jordan, and Iraq. This is a business that wins on the value of each order, not the number of new mouths. And that is exactly why it out-earns India on margin. A Gulf order is worth several times an Indian one while costing about the same to carry to the door, so more of each order survives as profit. That is why its food-delivery margin sits just above Zomato’s, even though Zomato serves a market more than a hundred times larger.

Who Competes With Talabat in the UAE, and Who Owns Them Now?

For years the UAE was a race between a few familiar names, and Talabat led it. On the most-cited panel it holds about 42 percent of the market by order value. Behind it sit Deliveroo, the British premium brand, at around 32 percent; Careem, the Gulf super-app, at about 18; and Noon Food, backed by a local billionaire and Saudi Arabia’s sovereign wealth fund, at about 8. The shares come from the same unaudited industry panel as the market estimate above; read them as the market’s shape, not a precise scoreboard.

Two of those names should make you look twice, because of who owns them now.

Deliveroo, the number two, was bought by DoorDash in 2025. Careem, the number three, belongs to Uber, the same Uber that is now buying Talabat’s parent. Uber bought Careem in 2019, ran the super-app arm that houses Careem’s food business as a joint venture with e&, the Abu Dhabi telecom group, and then took back control of that arm on the 30th of July 2026, two weeks after agreeing to buy Delivery Hero. So if the bigger deal closes, one company would sit behind both Careem and Talabat, roughly 60 percent of the UAE market on that panel. That is either the single biggest prize in the thesis or its single biggest regulatory problem, and it is why Gulf competition authorities are looking hard at the deal.

Then there is the newcomer that does not play by the old rules.

In late 2025 a fifth player arrived: Keeta, the international arm of Meituan, the Chinese delivery giant. It entered the way Chinese platforms enter, with a war chest and a tolerance for losing money. Free delivery, half off the first order, low commissions for restaurants, and no demand that they sign up exclusively. It reached Dubai in September 2025 and the rest of the country by the end of the year.

I saw what that looks like on actual bills, and, more useful, on bills six months apart. Five Keeta orders from the same Dubai account run from November 2025 to April 2026. The three November orders, placed within weeks of the launch, carry vouchers of 20, 35, and 50 percent off the food, and delivery is free on two of them; the third paid 1.90 dirhams. On one of those orders Keeta paid half the bill. The 2026 orders look different. In March, a 101-dirham order got no voucher at all, though delivery was still free. In April, a 14-dirham order got no voucher either, plus a forced 1-dirham top-up to reach the order minimum, plus 8.30 dirhams of delivery and service fees. The fees came to more than half the price of the meal.

I want to be careful with this, because five orders from one account prove nothing on their own. The November vouchers were probably partly new-customer offers, and those run out for any user. The April delivery fee may just be a small-order rule, since there is no 2026 order in this sample at the size of the November ones to settle it. What the bills do show, cleanly, is that the free-delivery-and-big-voucher experience this account got at launch is not the experience it gets now. If Keeta is starting to charge like a business rather than a land grab, this is what that would look like from the customer’s side.

The reason to take Keeta seriously is not the launch offers. It is what the same playbook already did next door. In Saudi Arabia, by one industry count, Keeta went from nothing to roughly the number two position, about a third of orders, in around fourteen months, by out-discounting rather than out-building. Other counts put its share lower, but none dispute the speed or the direction. It did most of its damage to the smaller players first, while the two leaders held on through sheer scale and bigger baskets. That is the live question hanging over Talabat’s richest markets. Not whether it keeps its lead, but whether Keeta pushes the cost of competing high enough to dent the margins that make this business special in the first place. Its funding is real but not endless. Meituan is losing money in a price war back home, and the thinning vouchers on these bills may be one small, local sign of that pressure.

How Do Delivery Apps Fight for UAE Customers? Subscriptions, Super-Apps, and Bank-Funded Discounts

Keeta’s vouchers are one part of a wider fight, and I did not have to guess at the rest of it. The friend’s(Dubai) account that this dive’s orders ran through also collects the competitors’ marketing, so I could sit and read how each player actually pitches a UAE customer.

What struck me, reading the pile, is that none of them is really pitching food delivery. Noon folds its food arm into noon One, a single subscription across everything the group runs: marketplace shopping, groceries, quick delivery, home services, free delivery on all of it. That is a direct answer to talabat pro, from the player with the deepest local backing. Careem, Uber’s super-app, pushes its dine-in arm with restaurant discounts of 15 to 25 percent, and it brings banks into the fight: one Al Hilal Bank card promotion ran 20 percent off Careem’s food, groceries, shopping, and dine-in bookings in a single month.

Then there is a layer under the platforms that I had not thought of as part of a delivery war until I saw it in the inbox: payments. Tabby, a buy-now-pay-later card (you split the bill into instalments), stacks its own money on top of whatever the platforms offer. Fifty dirhams off a Careem restaurant bill. A grocery programme whose partner list includes talabat mart and noon Food side by side, which tells you Tabby does not care who wins the delivery race. It is paying to own the payment underneath all of them.

For a Talabat thesis this changes two things. The everyday-app strategy Talabat is spending its margin on is not a private insight; Noon and Careem are chasing the same prize, which tells you how valuable it is and how crowded and the discount a UAE customer sees is not all platform money. Some of it comes from banks and card companies chasing payment volume, and a price war that other people help fund can run longer than any single platform’s losses suggest.

Is Gulf Food Delivery Regulated? Commission Caps, Exclusivity, and the UAE’s Middle Path

There is a common worry about this business: that governments across the Gulf are about to cap delivery commissions and crush the margins. It is worth checking against what the regulations actually say, because a cap on commissions is the one regulatory move that could break the economics I just described.

Here is what has actually happened, moving across the Gulf.

Saudi Arabia went first, in early 2025, with rules against predatory pricing, aimed squarely at the discounter cutting delivery fees to buy share. Qatar has been the enforcer. In 2025 it fined Talabat about 313,000 dollars and briefly suspended it for misrepresenting prices, and in 2026 it issued a formal code: every commission and fee disclosed in writing, pricing registered with the ministry. Years earlier it had even capped commissions outright after restaurant complaints, so the region’s platforms know Qatar will act. The UAE landed in the middle. It limited how long a restaurant can be locked into an exclusive deal to twelve months, and it made platforms disclose their commissions and how their apps rank restaurants. What the UAE did not do is cap the commission itself. The one hard cap in the Gulf today is Kuwait’s: in early 2026 it fixed delivery commissions and fees outright for three years. It is the outlier, and it is a smaller market.

So the honest read is calmer than the worry. The UAE has no commission cap. The regulation that does exist is mildly unhelpful to Talabat, because the thing being unwound, restaurant exclusivity, was one of the tools an incumbent uses to keep a discounter like Keeta out. But that is a long way from a margin-crushing cap. The pricing power that earns those six and a half cents on the dollar is, in Talabat’s biggest market, still intact.

Where Does Talabat’s Money Actually Come From, and Where Does It Go?

By now you have seen the app list: restaurant delivery, groceries in half an hour, a pharmacy, flowers, dine-in discounts, a paid membership, advertising space sold to restaurants. A list that long does not explain anything by itself. So strip it away, because underneath the list this company has one shape.

Talabat is a dominant, high-margin food-delivery engine in the Gulf, and everything else on the list is a use of that engine’s cash.

The engine first. In 2025, about 9.4 billion dollars of orders flowed across Talabat’s platforms, and 82 percent of that value came from the five Gulf countries. Of every dollar ordered, about 39 cents became Talabat’s revenue (that figure includes the groceries it sells directly, so it is not a pure commission rate), and 6.5 cents survived as operating profit. The engine carries no bank debt and sat on about 808 million dollars of net cash at the end of June 2026. That is the machine. The interesting part is where the money goes next.

A clarification before anything gets built on those numbers, because this piece says food delivery a lot. The 6.5 cents is the whole company’s profit rate, groceries and every other category included. Talabat reports its results by geography, not by business line, so nobody outside the company knows what the grocery arm alone earns or loses. Two things follow from that. Groceries and retail are already large: about a third of everything ordered on the platform by the end of 2025, by value and the company has said plainly that the build costs money right now, a cost the blended figure already absorbs, so the mature food business on its own earns more than those 6.5 cents, not less.

You can see the engine’s take on a customer bill first. The two orders I placed in August each came with a separate fee invoice: a delivery fee, a service fee, and VAT, adding up to AED 7.70 of fees one day and 6.95 the next. Both were charged even though the pro subscription was active on the account. To be fair to the product, talabat pro’s free delivery applies to eligible orders: above a minimum basket, from participating restaurants. These two evidently did not qualify. The point is not that the benefit is fake. It is that even a subscriber’s order can still carry seven dirhams of fees, and the engine collects them. And the customer’s fees are only part of the take. The commission the restaurant pays on the same order never appears on a customer invoice at all, because it is billed to the restaurant. Between the customer’s fees and the restaurant’s commission sits most of that 39 cents. Put these invoices next to the Keeta bills from the industry section and the difference in posture is the whole competitive story: the challenger opened by paying you to order. The incumbent charges you, twice, and still gets the order.

The first use is the dividend. On 2025’s profit, 421 million dollars went out to shareholders, 90.7 percent of what the company earned, and the stated policy is to keep paying out 90 percent. For a small shareholder, that payment, the roughly 5 percent a year from the summary at the top, is the visible return while everything else plays out.

The second use is the build. Talabat is spending part of the engine’s margin to become the everyday app you read about in the company’s story. talabat mart and its roughly 160 dark stores, the groceries, the pharmacy, the dine-in arm, the advertising business: each exists to raise how often you open the app. This build is exactly why the 2026 profit margin is guided down to roughly 4.7 to 5 percent of order value, from 6.5: management called 2026 a deliberate investment year of about 120 million dollars when it set the guidance in February. The profit decline you see in the 2026 numbers is that spending showing up in the accounts. Second-quarter profit fell 18 percent from a year earlier while order value grew 12 percent, and the gap between those two numbers is, mostly, the price of the build. On the company’s own disclosed measures, the build is working: three quarters of order value now comes from customers who use more than one vertical (a vertical is just a category: food, groceries, pharmacy), and over half comes from paying subscribers.

For scale on the ambition, I had Claude Code build a category-by-category grid of what Talabat, DoorDash, Delivery Hero, and the big Indian platforms each sell, from their own disclosures. Talabat’s list came out the broadest. It runs dine-in discounts and flowers that DoorDash does not, and it matches the Indian platforms category for category. The same grid also shows the gap: Blinkit, the grocery arm of the Indian group Eternal(Zomato parent) from the margin comparison earlier, runs around 2,400 dark stores to talabat mart’s 160. Talabat has the widest menu of services; it does not yet have the deepest operation behind each one. On quick groceries in particular, the Indian platforms are years further into the build. Their accounts also show what the build costs while it runs. Even at fifteen times talabat mart’s store count, the Indian group’s grocery arm still earns a fraction of what its food-delivery arm earns. A young grocery operation loses money for years before it makes any, and Talabat’s 2026 investment year is that same cost showing up earlier in the build.

The third use is the growth tail. Egypt, Jordan, and Iraq are about 22 percent of revenue and grew 46 percent in the first half of 2026. The Gulf engine barely grows in volume, around 5 percent a year, because nearly everyone who would order already does. The tail is where new users still exist. And, for now, it is also where Keeta is not.

Here is what seeing the company as one engine clarifies about the fight in the last section. Everyone in the UAE is building some version of the everyday app. The difference is the funding. Talabat pays for its build out of the engine’s own profit and still hands 90 percent of earnings to shareholders. Noon’s build is funded by its billionaire and sovereign backers, Keeta’s by a parent losing money at home, Careem’s by Uber and a telecom group, Deliveroo’s by DoorDash. Talabat is the only one whose war chest is the business itself. That self-funding is the strength of the model. It is also the exposure, because the build and the dividend are both paid out of the same six and a half cents that Keeta’s vouchers are aimed at.

One last piece of texture from the research account, offered as exactly what it is: a single customer, not a survey. The friend whose Dubai account this dive’s orders ran through tried Keeta through its launch months and has since settled on Talabat as the default, keeping Keeta for the occasional big group order, where a discount on a large basket bites hardest. One person proves nothing. But it is the pattern the subscription logic predicts: a discount wins single orders, while a membership decides where the everyday order goes without a thought. Talabat is betting the second is worth more, and on this one account, so far, it is right.

Why Is Talabat Stock Down 27 Percent From Its IPO Price?

Here is the uncomfortable arithmetic. Talabat sold at 1.60 dirhams a share in December 2024. It trades at 1.17 today, 27 percent lower, and in between the business grew its order value, grew its revenue, paid two dividends, and raised its 2026 guidance twice. A business this good does not usually sit this far below its offer price without reasons, and Talabat collected a full set: a rich starting price, a price war, a management change, a guidance cut, and an actual war. They arrived in that order.

The first reason is the price it started at. The offer was upsized on demand and priced at the top of its range, valuing the company at about 10.1 billion dollars, and the very first session told you how full that was: the stock closed its opening day at 1.49, 7 percent under the offer. It closed at 1.64 in January 2025, and that is still the high of its listed life. Anyone who bought in the first few months paid for a story that had no room to disappoint.

Then it disappointed twice, and neither time was really about demand. Between the end of May and the end of November 2025 the stock fell about 40 percent. The chart shows what walked in during those months. Keeta arrived in the Gulf: Qatar in August 2025, Kuwait in September, Dubai on the 27th of September, the rest of the UAE by the end of the year. The question it planted has hung over the stock ever since: what does a Chinese-funded price war do to the best margins in the industry? And in the middle of that, in October, the company said the chief executive who had run it since 2019 would hand over to a new one in November. Nothing broke in the numbers. The market just stopped paying a full price while it waited to see what Keeta and a new chief executive would mean.

The third blow was self-inflicted, and I covered its substance in the last section. In February 2026, off a year that delivered 615 million dollars of operating profit at a 6.5 percent margin, management guided 2026 profit down to 510 to 540 million, a margin near 4.7 to 5 percent, and called it a deliberate investment year. The market heard that margins were going down and sold first. By the 26th of February the stock had fallen 25 percent, to 0.715.

Then the region itself blew up. On the 28th of February the United States and Israel went to war with Iran, and this time the Gulf was not a spectator: Iranian missiles and drones struck targets on UAE soil, and the regulator shut the Dubai exchange entirely for two days in early March. When trading resumed, the whole market convulsed. Dubai’s index lost roughly 16 percent, around 45 billion dollars of value, inside a month. Talabat’s bottom, 0.658 on the 4th of March, 59 percent below the IPO price, printed in the middle of that panic. That low was not a judgment on the company alone. The guidance cut had already taken the stock from 0.956 to 0.715. The war panic pushed it the rest of the way, to 0.658, and that last leg had nothing to do with Talabat’s business.

Which also explains why the recovery came so fast. As the initial war shock faded from Gulf prices, a stock that had been hit twice got repriced twice. And the war itself, grim as it was, did not hurt this particular business: people who stay indoors still eat, and the half-year numbers Talabat delivered through those months came in ahead of its own plan. On the 12th of May the company raised its profit guidance with its first-quarter results, and the stock barely reacted; it actually slipped over the next session. Then the share buyback began executing on the 18th of May, and over the following two weeks the price rose about 30 percent. I am careful about drawing a straight line there, because the buyback’s actual purchases were small, about 35 million dollars in total, less than half a percent of the company. Whatever mix of fading war fear, bargain hunting, and improving numbers did it, by the end of June the stock had recovered to 1.23, above where it stood before the guidance cut. Worth knowing all the same: that supportive buyback has been idle since mid July, with barely a tenth of its authorization used, on no stated schedule.

The last reason is the one that does not show up in any income statement, and unlike the war, it has not faded. On the 16th of July, Uber agreed to buy Delivery Hero. Talabat’s own operations changed not at all that day, and the stock closed flat. Over the next two weeks it drifted down 11 percent. Then on the 12th of August the company raised all five of its guidance metrics, and the stock rose 5 percent the following day on about three times its usual volume, back to 1.17. Read those three reactions together and you can hear what the market is actually worried about. Good operating news still moves this stock up. The prospect of a new controlling owner moved it down, quietly, for two weeks.

So the diagnosis is specific. The war discount has largely unwound. The operating scares, Keeta and the margin reset, are real, but they are known, priced, and partly answered by two guidance raises since. What the market has not made peace with sits outside the accounts: a global competitor about to become the controlling shareholder, a technology licence that runs through that controller, and a minority position with no guaranteed exit price. The discount is the market charging for those three things.

Read more

Say the word timeshare and most people picture the 1980s version. A windowless conference room. A free breakfast that costs you a morning of your holiday. A salesman who keeps talking while a couple edges toward the door.

That version was real. In December 2016, Arizona’s attorney general announced an $800,000 settlement with a timeshare company called Diamond Resorts, after more than 500 complaints about what its salespeople had promised families in those rooms. The industry’s reputation got bad enough that a second industry grew up next to it: timeshare exit companies, which do nothing but help people get out of contracts they regret signing.

Hold that picture. Here is the same business in 2026.

Hilton Grand Vacations sells vacation packages from kiosks inside Bass Pro Shops and Cabela’s stores. 142 of them at last count, under a ten-year exclusive deal. You walk in for fishing gear, you walk out with a discounted trip to Orlando, and the trip includes a sales presentation.

The pitch also reaches you through Hilton Honors, the loyalty program you use for ordinary hotel stays. HGV’s license from Hilton requires it to take part in the program.

The product changed too. It is not a fixed week in a fixed room anymore. It is points in a members club, spendable across resorts and seasons. The newest tower is going up in Waikiki right now, and HGV was already selling it before the building was finished.

The buyer changed most of all. In the industry’s own 2026 owners survey, Millennials and Gen Z made 76% of recent purchases. The product your parents were warned about is now bought mostly by their children.

Hilton Grand Vacations, spun off from Hilton in January 2017, is the company this dive is about. It sells the new version of the product under the Hilton name. It also owns the old version: in 2021 it bought Diamond Resorts, the same company from that Arizona settlement. Both now sit inside one company, with about 722,000 members.

By HGV’s own numbers, the selling machine works. Tour flow, the count of families who actually sit through a sales presentation, has grown for four straight quarters. Tours are the raw material of this business. Get families into the room, and a predictable share of them buys.

On July 30, HGV reported its second quarter. Tour flow grew again. The sales those tours produced fell 2.9%. The stock fell 9.9% in one day.

Put those numbers next to each other. More families are touring than at any point in the last year, and they are buying less once they are in the room. Either something broke in how HGV sells, and that is fixable. Or the customers walking in can no longer afford to say yes, and that is a much bigger problem. The 9.9% fall says the market fears the second answer.

This dive is about finding out which one is true.

The 60-Second Thesis on Hilton Grand Vacations (HGV)

What they do. Hilton Grand Vacations sells vacation ownership, the modern form of the timeshare. A member pays upfront for points (a reusable currency for booking resort stays), then pays maintenance fees every year after that. HGV sells the memberships, manages the resorts and clubs for fees, and lends most of its customers the purchase money: 67% of buyers finance through HGV itself. So this is two businesses in one: a vacation seller with the Hilton name on the door, and a consumer lender with a $5.0 billion loan book.

Why the stock is down. On July 30, HGV reported its second quarter. Contract sales fell 2.9%. VPG, volume per guest (revenue per sales tour, the industry’s productivity number), fell 8.6%. The stock fell 9.9% in one day and sits at $45.66 as I write. The bigger problem is in the loan book. Loans more than 120 days late are 14.0% of the book, up from 12.6% six months earlier. The quarterly charge for expected loan losses reached $122 million, up 28% in a year, while sales shrank. The two closest peers, Marriott Vacations Worldwide (VAC) and Travel + Leisure (TNL), reported improving credit numbers over the same period. The deterioration is specific to HGV’s book.

Why it might be cheap. The share count. HGV bought back 35% of its own shares between the end of 2021 and July 2026, from 119.9 million to 77.7 million. Management raised its full-year EBITDA guidance in April and held it in July despite the weak quarter. Tour flow grew 6.1%, the fourth straight quarter of growth. By my math, the market prices the whole company, debt included, at about 6.6 times this year’s guided EBITDA (operating earnings before interest, taxes, and depreciation). VAC trades near 9.7 times and TNL near 8.3 times.

What could turn it. Three dated things. First, a new buyback authorization: $103 million remained as of July 23, and at the roughly $150 million a quarter pace of the past year that is about two more months, so the Q3 print is the test. The board has re-upped every year since 2022. Second, the Q3 credit numbers: if delinquencies flatten while peers hold steady, the bear case loses its best evidence. Third, Hawaii: sales at the unfinished Waikiki tower are deferred (booked as revenue only once the building is complete), so rooms already sold flip into reported earnings as construction finishes.

One honest caveat before we go further. If the loan book keeps deteriorating, the low multiple is deserved. The rest of this dive works that question out.

How Hilton Grand Vacations Was Built: the 2017 Hilton Spin-Off and Three Acquisitions

On January 3, 2017, Hilton split itself into three listed companies. Hilton kept the hotel brands. Park Hotels & Resorts took the owned hotel real estate. And the timeshare business became Hilton Grand Vacations, its own stock, its own board, ticker HGV.

Hilton kept zero ownership. What it kept instead is a license. HGV pays Hilton 5% of its gross revenues for the right to sell under the Hilton name, a fee that reached $192 million in 2025. The exclusive right to the name runs to 2051. And the license has teeth: HGV cannot even go through a change of control without Hilton’s consent. So HGV is independent, but only up to a point. The brand that makes its product sellable is rented.

Here is what the customer actually buys. A member pays upfront for points, then uses those points to book stays across the network, this year and every year after. On top of the purchase price come annual club dues and maintenance fees. And since two thirds of buyers finance the purchase through HGV, there is usually a third payment stream: interest, at an average rate of 14.4% across HGV’s loan book. That is one customer paying three ways: the purchase, the fees, and the interest.

Then HGV went shopping, three deals in five years.

August 2021: Diamond Resorts. Paid for with 33.9 million newly issued HGV shares. This deal did two things that matter to this story. It moved the old version of the industry, including the company from that Arizona settlement, inside HGV. And it handed the seller, the private equity firm Apollo, just over 30 million HGV shares, making it the largest shareholder. Hold the Apollo detail. It comes back when we get to why the stock is down.

January 2024: Bluegreen Vacations. All cash this time, about $1.6 billion including debt, at $75.00 per share, not one new HGV share issued. This is where the Bass Pro channel comes from: HGV signed the ten-year kiosk agreement in November 2023 while the deal was underway, and the acquisition brought a joint venture with Bass Pro that runs four wilderness resorts under the Big Cedar Lodge brand.

April 2026: Elara. HGV bought the 75% of its Elara joint venture in Las Vegas that it did not own from its partner Blackstone, for $131 million. Along with the tower came $411 million of customer loans and $310 million of debt attached to them.

Add it up and today’s HGV is the product of a five-year consolidation: about 722,000 members, $5.0 billion of revenue in 2025, and enough resort inventory to support $14.7 billion of future sales at current prices, by the company’s own measure. The industry spent four decades earning a bad name. HGV spent the last five years buying most of it.

The US Timeshare Industry in 2026: a $10.7 Billion Business With a Reputation Problem

Start with the numbers the industry publishes about itself. ARDA, the timeshare trade association, puts US timeshare sales at $10.7 billion for 2025, higher every year since 2022. About 10 million American households own one. And the resorts run full: close to 80% occupancy for two straight years, against roughly 62% for US hotels. Read those numbers knowing who published them, a lobby reporting on its own industry. But the direction matches what the public companies report: people keep buying this product.

The money is made in four layers, and they behave differently. A developer builds or buys resort inventory. Sales teams sell it as points or weeks, and selling is the expensive part of the model: the discounted packages, the kiosks, and the tours all sit in the cost line. The financing layer lends buyers most of the purchase price at mid-teens interest rates. And the management layer collects club dues and maintenance fees every year, under contracts that renew. That last layer is the closest thing this industry has to subscription revenue.

The product also has a structural reputation problem, and it shows up in a strange place: the escape route. Getting out of a timeshare contract is hard enough that a whole industry of exit companies exists to sell escape. That industry then produced its own scandal. Washington State’s attorney general sued the largest exit firm, Reed Hein, known as Timeshare Exit Team, in February 2020. The September 2021 settlement cost the firm $2.61 million plus a $22 million suspended judgment (a penalty collected only if the settlement terms are broken), mandatory customer refunds, and a public apology. Even leaving this product safely required a regulator.

The modern industry consolidated behind hotel names. Marriott spun off its timeshare arm in 2011; that company, Marriott Vacations Worldwide (VAC), bought ILG, another timeshare and exchange operator, for $4.6 billion in 2018. Wyndham split itself in 2018 too, and its timeshare half bought the Travel + Leisure brand in 2021 and took the name, Travel + Leisure Co. (TNL). Hilton’s arm became HGV in 2017, then bought Diamond and Bluegreen. Three hotel-heritage companies now front the US industry. That is not an accident: after decades of reputation damage, the industry’s fix was to sell under hotel brands people already trust. Remember VAC and TNL. They are the two comparison companies this dive keeps returning to.

Where the cycle sits in mid-2026: mixed. US air passenger volumes fell 1.3% year over year in July, a second straight monthly decline, per the US Travel Association. Resort hotels grew RevPAR (revenue per available room, the hotel industry’s core sales measure) by 4.6% against 13.7% for urban hotels, and the association notes that travel spending growth is now mostly higher prices, not more travelers. PwC projects 2026 hotel RevPAR up 2.9%. Inside timeshare, ARDA’s own reports show the average transaction price fell to $23,160 in 2024 from $24,170 in 2023. The customer is still traveling and still paying more. Volumes are flat to down; prices are doing the work.

One fact from this section matters more than the rest. Industry sales rose in 2023, 2024, and 2025. HGV’s contract sales fell 2.9% in the June quarter. A company shrinking inside a growing industry is losing share. Keep that in mind when management explains the weak quarter as an execution problem. We get there in three sections.

How HGV Actually Makes Money: a Vacation Seller With a Lender Behind It

The way to see this company clearly is to split it in two.

Company one sells vacations. It builds or buys the towers, runs the kiosks and the tours, and closes the sale. This is the loud, expensive half. In 2025 the selling activity made $453 million of profit at a 21.2% margin, down from 24.1% a year earlier. Attached to it is a rental business that lost $39 million renting out unsold rooms, mostly because HGV pays the maintenance fees on inventory it has not sold yet.

Company two starts earning the moment a sale closes, and does not stop. It lends buyers the purchase money: $5.0 billion of loans outstanding at a 14.4% average interest rate. And it manages the resorts and clubs for fees that come back every year. In 2025, financing made $298 million of profit at a 58% margin. Resort and club management made $551 million at 71%.

Add up each half. The after-sale streams made $849 million of profit in 2025. Selling, including that rental loss, made $414 million. Two thirds of HGV’s profit comes from financing and managing people who already own, not from selling to new ones (by my math, from the profit lines in the 10-K).

Two consequences follow, and they drive the rest of this dive.

First, the fee-and-interest machine only grows if the selling machine keeps feeding it new members. That feed has stopped: 721,896 members at June 30, slightly fewer than a year earlier. Growth right now comes from selling more to the people already inside, and 74% of 2025 contract sales came from existing owners.

Second, the profit engine is a lender, so it is only as healthy as its borrowers. HGV lends at 14.4% and funds those loans by borrowing against them at about 5% (and on most of that debt, lenders can only claim the loans themselves, not HGV’s other assets: it is non-recourse). That spread is the best business inside the company. But when borrowers stop paying, the expected loss is charged against sales revenue, and that charge reached $122 million in the June quarter alone. The lender’s health is the thread the rest of this dive pulls.

The balance sheet in one line: $4.9 billion of corporate debt at a 5.6% average rate, another $2.9 billion borrowed against the loan book without recourse to HGV, and about $272 million of cash.

And capital allocation in one line: nearly every spare dollar goes to buying back HGV’s own shares. That arithmetic is unusual enough to get the next section to itself.

Share Buybacks Explained: AutoZone, Bed Bath & Beyond, and HGV’s Disappearing Share Count

HGV has retired 35% of its own shares in four and a half years. Before deciding whether that is good news, it helps to know when a buyback creates value and when it destroys value, because the same act sits behind a stock that rose roughly 95x in 27 years and behind a company that went bankrupt.

The mechanics are simple. Picture the company as a pie, cut into one slice per share. In a buyback, the company uses its own cash to buy slices back and remove them from the pie plate. The pie itself stays the same size; it is just cut into fewer slices now, so every remaining slice is bigger. That is all a buyback does.

The common belief is that this automatically creates value. It does not. A buyback is an investment like any other, and it depends on the price paid: pay less for a slice than the slice is actually worth, and everyone still holding a slice comes out ahead. Pay more, and the company has overpaid for its own stock and made the remaining owners poorer.

The second thing that decides the outcome is where the cash comes from. Buy back stock with genuine spare cash, and the worst case is a wasted opportunity. Buy it with borrowed money while the business weakens, and the company reaches the downturn with the debt still there and the cash gone.

Here are both outcomes, from real filings.

AutoZone is the success case. From 1998 to 2025 the car-parts retailer cut its share count from 150.4 million to 16.6 million, an 89% reduction, spending about $38.6 billion on it (by my math: $39.2 billion authorized over the years, minus the $632 million still unused). Here is the important part. The business grew well but not spectacularly: net income rose about 11x over those 27 years. Earnings per share rose about 98x over the same period. The difference between 11x and 98x came entirely from the shrinking share count. The stock went from $32.94 at the end of 1998 to about $3,127 this month, roughly 95x, with no stock splits.

NVR did the same thing. The homebuilder cut its share count 62.6% between 2001 and 2025 and is still buying today ($1.82 billion of buybacks in 2025 alone). Its stock went from $204 to about $6,390 over that period, roughly 31x.

Bed Bath & Beyond did the same thing and went bankrupt. It repurchased $10.5 billion of its own stock from 2004 through fiscal 2017. Then its retail business declined, and the losses arrived after the cash had already been spent. In fiscal 2022 it lost $36.03 per share. To keep operating it had to sell new shares at very low prices, and by May 2023, two weeks after its Chapter 11 bankruptcy filing, the share count stood at 739 million, up from about 97 million a year earlier. The stock closed at $0.29 on the last Friday before the filing. The buybacks did not cause the decline of the business. They spent the $10.5 billion the company could have used to survive it.

So the outcome is not decided by the buyback itself. Three checks separate AutoZone from Bed Bath & Beyond: was the price paid below per-share value, was the cash genuinely spare, and was the underlying business growing or declining.

Run the checks on HGV. The scale is real: $272 million of buybacks in 2022, $368 million in 2023, $432 million in 2024, $600 million in 2025, and $325 million so far this year through July 23, at an average of $48.19 per share last quarter. By my math, each remaining HGV share now represents 54% more of the company than it did at the end of 2021.

But all three checks are still open, and the rest of this dive works through them. Is the price paid below what the business is worth per share? That is the valuation section. Is the cash genuinely spare? HGV’s own headline cash-flow measure counts money borrowed against its loan book as cash flow, and separating real cash from borrowed cash is also the valuation section’s job. Is the underlying business growing or weakening? That is the question the next two sections answer.

One more number before moving on. The buyback authorization has $103 million left. At the recent pace of roughly $150 million a quarter, that is about two more months of buying. The board has approved a new authorization every year since 2022. Whether it does so again at the Q3 report is the nearest-term test of the whole HGV story.

Why HGV Stock Fell 9.9%: the Weak Quarter, the Loan Book, and the Apollo Overhang

Four things are weighing on the stock. Here they are, one at a time.

The quarter itself. Contract sales (the total value of new vacation-ownership purchases signed in the quarter) came in at $810 million, down 2.9% from a year earlier. VPG (volume per guest: the average sales value each tour produces) fell 8.6% to $3,372, while tour flow (the number of people who sat through a sales presentation) rose 6.1%, the fourth straight quarterly increase. On the earnings call, management gave two explanations. Bluegreen resorts were comparing against a quarter last year when buying jumped after the launch of HGV Max, the combined membership program. And a handful of Bluegreen sites in Orlando and Myrtle Beach had what management called execution challenges, where new sales leadership has been installed. Management also said it expects third-quarter contract sales to fall by a mid single digit percentage. One more item sat in the quarter: a $48 million loss on the disposal of certain resort properties. HGV has not disclosed the buyer or the terms of that sale.

The loan book. This is the part the market is most afraid of. Remember how this business works: HGV lends most of its customers the money for the purchase. Some of those customers will eventually stop paying, and accounting rules require HGV to estimate those future losses in advance, at the moment of sale, and subtract the estimate from revenue. That estimate is called the provision (the charge for expected loan losses).

The provision reached $122 million in the June quarter, up 28% from a year earlier, even though the sales that create the loans fell. Measured against contract sales, the provision was 15.1%, up from 11.4% a year before. In plain terms: for every $100 of vacations HGV sold in the quarter, it set aside about $15 for customers it expects to stop paying, up from about $11 a year ago.

The other loan-book numbers point the same way. Loans more than 120 days late grew to 14.0% of the loan book at June 30, from 12.6% six months earlier. The allowance (the total pool of money already set aside for expected losses on the loans HGV made itself) has risen every reporting period since the end of 2022, from 23.0% of those loans to 30.1%. And in the first half of 2026, HGV gave up on and wrote off $107 million of loans, about the same as in all of 2024.

One possible explanation is that American consumers are stretched everywhere. The filings of the two closest competitors do not support it. Over the same period, Marriott Vacations reported delinquencies (payments running late) improving by about one percentage point from a year earlier, and Travel + Leisure reported its share of past-due loans falling. Whatever is happening to HGV’s borrowers is not happening across the industry.

The Apollo overhang. Apollo, the private equity firm, received just over 30 million HGV shares when it sold Diamond Resorts to HGV in 2021, which made it the largest shareholder. It has been selling in blocks: 8.05 million shares at $42.85 in August 2025, then 5.75 million at $50.00 in June 2026. HGV used its own buyback to purchase part of each block, 933,488 shares the first time and 750,000 the second. Apollo still holds about 12.5 million shares, 15.9% of the company, and HGV’s own annual report now states that Apollo may continue to sell. Buyers know more shares are likely to be sold, and that knowledge holds the price down. Two related details: one of Apollo’s two board seats went away on July 2 when its designee resigned under the shareholders agreement, and if Apollo sells about another 528,000 shares, its board representation ends entirely.

The insider tape. Since the July 30 report, no HGV insider has bought shares in the open market. One has sold: the general counsel sold 20,691 shares on August 6 at an average price of $46.90, about $970,000, roughly 30% of the shares held directly before the sale (unvested stock awards sit on top of that, so the share of the total stake is smaller). The filing shows the sale was not made under a pre-arranged trading plan (a 10b5-1 plan, which schedules insider sales in advance). One sale by one executive proves nothing; people sell for taxes and personal reasons. The plain fact is this: a week after a 9.9% fall, the insiders’ combined open-market activity was one sale and zero buys.

Weigh the four parts. The Apollo shares will be sold and absorbed eventually. A badly run quarter can be fixed with better selling. The loan book is the part that decides everything, because if HGV’s customers are running out of ability to pay, then the falling VPG is not an execution problem, the loss charge keeps growing, and the earnings that fund the buyback shrink. The whole thesis now hangs on one question: is this a selling problem or a customer problem? The next section takes both sides of it seriously.

The One Question That Decides HGV: a Selling Problem or a Customer Problem?

This whole dive comes down to one question. The number of people taking HGV’s sales tours is growing. The amount those people buy is falling. Why?

There are two possible answers. Answer one: HGV’s sales teams are doing a bad job of selling. That can be fixed within a few quarters. Answer two: HGV’s customers can no longer afford the product. That is a much bigger problem. Below is the strongest honest case for each answer, and then my own conclusion.

The case that the customer is the problem.

The strongest evidence is in the loan numbers. In April 2026, HGV acquired Elara, and Elara’s customer loans came with it. So you could ask: maybe the loan book only looks worse because of these newly added loans? The filings answer that question. Remove the Elara loans completely and look only at the loans HGV already had: loans more than 120 days late still rose from 12.6% of that book to 14.1% in six months. So it is not the newly acquired customers making the numbers worse. HGV’s existing borrowers are paying worse than they did six months ago.

The second piece of evidence is history. HGV’s annual reports disclose what percentage of its loans defaulted each year (a default means the customer stopped paying and the loan was declared a loss). In 2019, before the pandemic, that rate was 5.14%. In 2024 it was 10.77%. In 2025 it was 9.86%. Defaults today run at about twice the pre-pandemic level. And HGV did not loosen its lending rules to cause this: the filings have described the same minimum 10% down payment since 2021. The rules for getting a loan stayed the same. The customers passing those rules got weaker. One more sign points the same way: the average new loan has been getting longer, from 8.2 years remaining in 2022 to 8.9 years in 2025. Lenders usually stretch loans out longer when buyers need smaller monthly payments.

The third piece of evidence is about future tours. Customers pay in advance for discounted vacation packages, and those packages are what bring them to sales presentations later. So the balance of these advance payments tells you how many future tours are already paid for. That balance was $226 million at the end of 2024 and $228 million at June 30, 2026. It has not grown in a year and a half, even though tours grew 6.1%. In other words, HGV is using up its supply of already-sold packages, and it is not selling enough new ones to replace them.

Now add what the last section showed. The two closest competitors reported improving loan numbers over the same period. And no HGV insider has bought a single share since the stock fell on July 30. If this case is right, then the 8.6% fall in VPG is not a sales-team problem. It means the customers themselves are running out of money, and next year’s earnings will come in lower than this year’s, whatever the guidance says today.

The case that the selling is the problem.

The first fact: tours keep growing. Tour flow has now grown for four straight quarters. Families are still buying the discounted packages, still getting on planes, still sitting through the presentations. People who could not afford the product would eventually stop showing up. They have not stopped.

The second fact: last year’s comparison was unusually hard to beat. On the earnings call, management explained that Bluegreen resorts were being measured against the June 2025 quarter, when buying per tour jumped about 45% after the launch of HGV Max, the combined membership program. When you compare a normal quarter against an unusually strong one, the normal quarter looks weak even if nothing is wrong. Management also named the exact places where selling went badly, Orlando and Myrtle Beach, and replaced the sales leadership there. That explanation is specific enough to be checked: within two quarters, either those markets recover or they do not.

The third fact: management did not cut its profit guidance. The full-year guidance for Adjusted EBITDA (operating earnings before interest, taxes, and depreciation, on the company’s own adjusted definition) was raised in April and kept unchanged in the same July release that reported the weak quarter. A management team watching its customers run out of money would normally cut that number. This one did not.

On the loan numbers, this case has three answers. First, high defaults are normal in this industry. Marriott Vacations’ own filings assume that 13.05% of its loan balances will eventually default. Travel + Leisure sets aside 20.8% of its sales for expected losses. HGV’s loss levels are inside the industry’s normal range. What is unusual at HGV is the direction of change, not the level. Second, the provision is an estimate made in advance, not cash actually leaving the company that quarter. HGV books the full expected lifetime loss on a loan on the day of the sale. So a cautious estimate lowers reported profit today, while the real cash losses arrive slowly, over years. Third, Fitch and trade press reports through 2025 said that timeshare loans were performing worse across the whole industry, especially loans made in 2025. So part of HGV’s bad trend may be an industry problem, not an HGV problem.

And one thing makes the reported quarter look worse than the real quarter: Hawaii. Sales at the unfinished Waikiki tower are deferred (booked as revenue only once the building is complete). So the reported earnings exclude selling that has already happened.

If this case is right, then the weak quarter was a hard comparison plus two badly run sales markets. The guidance holds, the Hawaii sales get added to reported earnings as construction finishes, and the buyback keeps retiring shares in a business the market just marked down 9.9%.

Before I give you my read, make your own call.

My read. The two cases are actually answering two different questions. Once you see that, most of the confusion clears up.

Question one: why did VPG fall 8.6% this quarter? Here the selling explanation mostly convinces me. The Bluegreen comparison problem was real, the problem markets are named, and the fix can be checked within two quarters.

Question two: why does the loan book keep getting worse? Here the selling explanation cannot help at all. How well a sales team performs this quarter has no effect on whether people who bought in 2023 keep paying their loans in 2026. And the loan book is getting worse on its own: excluding Elara, late loans rose from 12.6% to 14.1% in six months, while both competitors improved.

So my answer is split, and I mean the split precisely: the weak quarter is mostly a selling problem, but the loan book shows a real weakening of HGV’s financially weakest customers, and that part is specific to HGV. The loan book matters more for the stock. Worsening loans feed the provision, the provision reduces profit, and profit is what pays for the buyback.

Here is what would prove me wrong. I am writing it down now, with dates, so that neither you nor I can reinterpret it later. Management itself guided third-quarter contract sales down by a mid single digit percentage. If the actual fall is worse than that, or if VPG falls anywhere near 8% again after the Bluegreen comparison has passed, then the selling explanation is dead and the customer explanation wins both questions. And if the third-quarter loan numbers worsen again while Marriott Vacations and Travel + Leisure hold steady, my split answer gets worse, not better.

One more thing about the loan book belongs in this dive before the valuation. The cleanest loan numbers HGV shows the market are not what they appear to be. That is the next section.

Why HGV’s Loan Bundles Look Perfect: the Company Buys the Bad Loans Back at Full Price

One argument comes up in every bullish discussion of HGV, and it sounds strong. It goes like this: if HGV’s customer loans were really going bad, the bond market would notice, because the bond market lends against those exact loans. And the bond market keeps lending to HGV, cheaply and eagerly. In June 2026, HGV sold a $300 million bundle of customer loans to bond investors. Investors wanted about nine times more than was available, and HGV’s CFO said the deal priced at the tightest spread (the extra interest cost above benchmark rates) in timeshare since January 2022. How can the bond market be that calm if the loans are going bad?

That argument is wrong, and the proof is in HGV’s own filings. To understand the proof, you first need to see how HGV turns its customer loans into cash. Here is the whole process, step by step.

HGV does not keep its customer loans and wait ten years to collect the monthly payments. Instead, it does what banks do with home mortgages. It gathers thousands of customer loans into one large bundle. Then it sells that bundle to bond investors. The investors pay HGV cash immediately, about 98 cents for every dollar of loans in the bundle. In exchange, the investors receive the customers’ monthly payments as those payments come in. This is called securitization (turning a pool of loans into bonds that investors can buy), and it is normal practice across finance. Home mortgages, car loans, and credit card balances are funded the same way. Marriott Vacations and Travel + Leisure fund their timeshare loans the same way too.

Why does HGV do this? Because it gets its cash back on day one instead of over ten years, and it can immediately lend that cash to the next customer. The bond investors are, in effect, funding HGV’s lending business year after year. HGV needs them to stay comfortable.

With that background, look at what HGV’s latest quarterly filing discloses. The filing reports how many loans are more than 120 days late, and it reports this separately for loans inside the bundles and loans outside them. Of the loans HGV made itself and has NOT put into bundles, 25.2% are more than 120 days late. Of the loans HGV made itself that sit INSIDE the bundles sold to bond investors, 0.7% are more than 120 days late. Same company, same product, same kind of customer: 25.2% late outside the bundles, 0.7% late inside them.

I found this split by having Claude Code read the loan footnotes of every HGV filing back to 2021 and separate every late-loan number into two piles: loans inside the bundles and loans outside them. The 0.7% against 25.2% gap is what came out of that sorting.

Part of the gap is screening: only healthy loans are allowed into a bundle in the first place. But loans go bad after they enter bundles too, and the bundles still show 0.7%. How? HGV’s own filings give the answer, in one sentence: “we often replace or repurchase timeshare financing receivables that are in default at their outstanding principal amounts.” In plain English: when a loan inside a bundle goes bad, HGV often buys it back at full face value (at par, meaning 100 cents on the dollar), or swaps a healthy loan in. The bad loan moves back onto HGV’s own books. The bundle stays clean, the bond investors stay happy, and the bad loan joins the group of loans outside the bundles. That is part of why the outside group shows 25.2% late.

How much does this support cost? The filings never say. Nowhere does HGV disclose how many dollars of defaulted loans it buys back from the bundles each year. The cost sits inside the overall loan-loss numbers you have already seen, without its own line. The only outside estimate comes from Fitch, the rating agency that grades these bonds. According to Fitch’s surveillance reports, on HGV’s older bundles from 2018 through 2020, defaults equal to roughly 2.4% to 3.5% of the original bundle size were bought out or swapped out by HGV. Fitch also states the consequence: none of these deals has ever produced a loss for bond investors, and Fitch credits the repurchases and substitutions for that record. The bond investors have never lost a dollar on HGV’s loan bundles, partly because HGV keeps paying to make sure they never do.

Two more outside observations point the same way. In October 2025, Fitch raised its lifetime default expectations on HGV’s 2024 bundles, meaning those loans are performing worse than Fitch originally assumed. And the average credit score (FICO, the standard US credit score) of the loans in HGV’s newest bundle, sold in December 2025, was 727, which Moody’s called one of the lowest in the history of HGV’s deals, down from 745 in a bundle sold in early 2025.

So the bullish argument gets the direction wrong. The bond market’s calm is not independent proof that HGV’s loans are healthy. The bond market stays calm partly because HGV spends its own money to keep the bundles clean.

Three disclosure changes, all in the same direction. While the loan numbers worsened, three things also changed in what HGV shows the public.

First, the default-rate table. HGV’s quarterly reports used to include the annual default rates you saw in the last section. Both 2026 quarterly reports dropped that table. The numbers still appear once a year in the annual report, but the quarterly view of defaults is gone.

Second, the credit-score method. In the first quarter of 2026, HGV changed how it calculates the borrower credit scores it discloses, from an average of the available scores to a single score. The new method produces higher numbers. On the old method, the disclosed score had peaked at 741 in 2024 and fallen to 734 in 2025. The new method shows numbers around 747 to 751. Even on the new method, the score fell from 751 to 747 year over year.

Third, the Manhattan Club case. HGV inherited an arbitration case (a private court proceeding) with its Bluegreen acquisition, and in 2024 the arbitration panel ruled against HGV’s subsidiary. The 2025 annual report describes what HGV had to do: take on a $47.5 million note payable and make forced inventory purchases of $7.5 million every quarter, with the damages phase still unresolved. Then both 2026 quarterly reports stopped mentioning the case entirely, while the accrued legal liabilities line went from $8 million to $19 million in a single quarter, with no explanation given.

I want to be careful here. There may be an innocent explanation for each change on its own. Companies drop tables they decide are immaterial. A methodology change can be a genuine improvement. Legal disclosures come and go as lawyers judge what is material. And I am not claiming anyone at HGV intends to hide anything. I cannot know that, and the filings do not say it. What I can say is simpler: three disclosures changed in the same year, all three in the direction of showing less about a loan book that was getting worse. I weigh that as a pattern.

The catch. The repurchases themselves are not a scandal. They are optional, they are disclosed (in that one sentence), and they may simply be good business: the bundling channel gives HGV cheap funding at 98 cents lent per dollar of loans, and protecting that channel is worth real money. A rational finance chief might make the same choice with clean motives. The same sentence has also appeared in HGV’s filings since at least 2021, so this is not new behavior invented to hide this year’s problems. The reason it matters is narrower: you cannot use the bond market’s calm as proof that the loan book is fine. The loan book’s own numbers, the ones you saw in the last two sections, are the evidence that counts, and they are worse than the clean bundles suggest.

Two questions I could not answer from the filings. How many dollars per year do the repurchases consume? And when a customer who is behind on payments upgrades to a bigger package, which replaces the old loan with a new, larger one and resets its age, are late-paying customers allowed to do that? The filings do not say. I am sending both questions to HGV’s investor relations. If I get a reply, it will appear here as an edit at the end of this piece, or in a future update.

Now the valuation. Everything so far goes into it: the two businesses inside HGV, the buyback, the split answer at the hinge, and a loan book whose cleanest numbers I now trust less.

What I Think Hilton Grand Vacations Is Worth: My Bear, Base, and Bull Cases

Before the numbers, three warnings. A valuation is a set of assumptions multiplied together. Mine is sensitive to three of them: how much HGV earns in 2027, how much of that turns into real cash, and what price the market will pay for that cash. Change any one of the three and the answer moves a lot. Everything below is my opinion, with the method shown step by step, so you can disagree with a specific step instead of just the conclusion.

The first job is deciding what HGV actually earns, because HGV’s profit comes in three versions. Version one is reported Adjusted EBITDA (operating earnings before interest, taxes, and depreciation, on the company’s own adjusted definition): $950 million in 2025 for the portion belonging to HGV shareholders. Version two adds back the deferred sales (the Hawaii-style sales already made at unfinished buildings, not yet allowed to be counted as revenue): $1,152 million. This second version is the one HGV’s own bonus plan pays managers on, so I treat it as the honest measure of what the business earned. Version three is the guidance number for 2026, $1.225 to $1.265 billion, also calculated before deferrals. Three versions of profit, about $300 million apart. I build everything on version two.

The second job is turning earnings into cash. Here my method differs from the company’s own headline number, and the difference matters a lot. HGV publishes a measure it calls Adjusted Free Cash Flow. For 2025, that measure was $756 million. Here is what was inside it, from HGV’s own reconciliation table: $154 million of actual free cash flow, $404 million of new net borrowings against the loan book, and about $198 million of added-back costs. More than half of HGV’s 2025 “free cash flow” was borrowed money. One quarter, the fourth of 2024, showed $883 million on this measure, and $747 million of that was borrowings. In the first quarter of 2026, the same measure was negative $37 million. The reason for these wild swings is simple: the number rises and falls with how much HGV borrows in each quarter. So it cannot tell you how much the business actually earns.

So I built my own cash number. Start with the $1,152 million of version-two earnings. Subtract the interest HGV pays on its corporate debt. Subtract cash taxes. Subtract what HGV spends on property, software, and new resort inventory. Subtract the cash tied up in funding new customer loans. Do not count any borrowed money as income. What is left is cash that actually belongs to the shareholders. On 2025’s numbers this figure is much smaller than the company’s headline, and much steadier. All three cases below are built on it.

One more concept before the cases, because the cases depend on it: the multiple. When I say a share is worth nine times its yearly cash, I mean the buyer is paying today for nine years’ worth of the cash that one share produces. A growing, safe business deserves a high multiple, because its future years will probably arrive and will probably be bigger. A shrinking or risky business deserves a low multiple, because its future years may never arrive. So in each case below, I first work out the cash per share, and then decide how many years of that cash the business deserves to be priced at.

Now the three cases, all valued at the end of 2027.

My bear case: about $26 per share, roughly 42% below today’s $45.66. In this case, 2027 earnings (version two) fall to $1.0 billion: the fee HGV pays Hilton steps up as scheduled, loan-loss charges keep climbing at the current pace, and demand weakens. My cash number comes out near $344 million. That includes extra cash taxes, because HGV currently postpones tax on installment sales (sales paid over time), and that postponement unwinds when the loan book stops growing. The buyback stops when the current authorization runs out, leaving about 72 million shares. Divide $344 million of cash by 72 million shares: each share produces about $4.80 of cash a year. I price that at 5.5 times, a low multiple for a business in trouble, and get roughly $26 per share. This is what I think the stock is worth if the customer explanation is true and HGV’s borrowers really are running out of money.

My base case: about $76, roughly 66% above today. Earnings reach $1,275 million, a little above the top of this year’s guidance range, as the Hawaii sales enter the numbers and credit stabilizes without improving. Cash comes out near $537 million. The buyback continues and the share count falls to about 63.9 million, so each share produces about $8.40 of cash a year. I price that at nine times and get roughly $76.

My bull case: about $103, roughly 126% above today. Earnings reach $1.4 billion: Hawaii lands, the two problem sales markets recover, and credit actually turns. Cash comes out near $593 million across about 63.3 million shares, which is about $9.40 of cash per share. I price that at eleven times and get roughly $103.

A second method as a sanity check. Value the whole company at seven times version-two earnings (this values the enterprise: the shares plus the debt together), subtract the roughly $4.6 billion of net corporate debt, and divide by the shares. The base case comes out near $70. The two methods land close together, which makes me trust the answer more.

Now put HGV next to its two peers, because that comparison does two jobs at once: it checks whether my multiples are reasonable, and it shows what the market already charges for the risk. At the August 7 closing prices, by my math, the market values HGV at about 6.6 times this year’s guided earnings. Travel + Leisure trades near 8.3 times its earnings. Marriott Vacations trades near 9.7 times. HGV is the cheapest of the three, by a wide margin.

The discount has a reason, and you already know it: HGV is the only one of the three whose loan book is getting worse. Marriott Vacations and Travel + Leisure also carry billions in customer loans and also expect double-digit defaults, but their numbers improved this year while HGV’s worsened. The market is charging HGV for that difference. My cases simply price the two outcomes: the base case assumes HGV closes part of the gap to Travel + Leisure, not all of it, and the bear case assumes the gap gets wider.

You will see higher numbers than mine elsewhere. Some investors put a multiple of ten directly on the company’s own Adjusted Free Cash Flow measure, which produces values well above $100 per share. I do not, for the reason above: that measure counts borrowed money as cash flow. On the cash the shareholders actually receive, my base case is the mid-$70s. If you use that higher math, you should at least know that the number underneath it includes borrowed money.

One more piece of arithmetic, a discipline I run on every company I research. I ask: at what price would the upside to my base case be exactly twice the downside to my bear case? With a base of $76 and a bear of $26, that price is about $42.7. Today’s $45.66 sits about 7% above the point where my numbers say the reward is twice the risk. Close, but not there. That is a fact about my arithmetic, not an instruction to anyone.

My full Excel model comes attached with this piece, the same way as in the previous dives: my inputs in blue, so you can change any assumption and watch the answer move.

View the valuation model

Those are my numbers. The next section lists the ways they could be wrong.

What Breaks the HGV Thesis: Five Risks and the Sign for Each

I tried to kill this thesis before publishing it. These are the five ways it dies, each with the sign that would tell me it is happening.

1. The board does not renew the buyback. About $103 million remains of the current authorization, roughly two more months at the recent pace. The board has approved a new authorization every year since 2022. If the Q3 report comes and goes without a new one, the shrinking share count that drives the whole per-share story stops. It would also be information in itself: boards pause buybacks when they see trouble coming or need the cash somewhere else. The sign: the Q3 report itself, and whether a new authorization comes with it.

2. The loan book keeps getting worse. Every quarter of rising late payments adds to the provision, and the provision comes straight out of profit. If the provision keeps climbing from the current 15.1% of contract sales, my base case earnings are wrong and the bear case is the honest map of where this goes. The sign: HGV’s Q3 credit numbers, read next to Marriott Vacations’ and Travel + Leisure’s.

3. A real consumer downturn arrives. Everything in this dive happened while US employment was healthy. HGV’s defaults already run at about twice their 2019 level in a decent economy. A recession hits the weakest borrowers first, and HGV’s problem is already concentrated in its weakest borrowers. A timeshare is also one of the first purchases a worried household delays. In that world, my bear case of $26 stops being the cautious scenario and becomes the likely one. The sign: US unemployment and consumer-credit data turning together.

4. The Hilton license terms turn against HGV. The fee HGV pays Hilton for the brand is stepping up on a schedule: the Diamond resorts reach the full 5% rate around August 2026 and the Bluegreen resorts by early 2028, a built-in cost increase of roughly $40 million a year by 2027, by my math. Two harder clauses sit behind it. If HGV misses its room-conversion targets badly enough, the fee escalates further, and Hilton gains the right to block future sales of HGV Max, the combined membership program. And any sale of HGV itself requires Hilton’s consent, which closes off the most common ending for a cheap stock, a takeover. The sign: any disclosure about missed conversion milestones.

5. The Manhattan Club case lands a big number. The damages phase of the arbitration is still unresolved, and HGV’s own history says these cases can be large: a different inherited case, from the Diamond acquisition, ended in early 2024 with a payment of about $104 million, of which HGV paid about $50 million in cash and insurance covered the rest. Meanwhile, accrued legal liabilities jumped from $8 million to $19 million in the June quarter with no explanation given. If the damages award lands anywhere near the earlier case’s scale, that is a direct cash hit in a year when only $103 million of buyback authorization remains. The sign: any 8-K or footnote resolving the damages phase.

That is the full list of what I think could break this thesis. Now the verdict.

Where I Land on HGV: My Verdict, Dated August 9, 2026

Time to land this.

Everything you have read so far is free, in full, on purpose. I would rather show you what a Research Dive is than describe one. What sits below is the part paid subscribers get with every dive: my model portfolio entry, its size, the reasoning behind that size, and the three tests that will decide what I do next.

Read more

$155,579,371 in cash.

$292,644 of interest income.

So both numbers from above sit in the same annual report of the same NASDAQ-listed company. The same got audited 10K, filed with the SEC, and then signed by management. Nothing in the filing says anything wrong

But park $155 million in plain time deposits and it earns about three million dollars a year. This company reported less than $300,000. That is a yield of 0.19 percent. Either management chose to leave nine figures earning almost nothing, or the cash was never there to earn anything.

So I ran a test. I stripped the company’s name and country out of the filing data and handed it to four AI analyst agents running blind. Four separate sessions, no shared memory, no hint that anything was wrong. (I ran them as Claude Code subagents. Claude Cowork, the desktop app that runs the same engine without a terminal, can split the job into parallel agents the same way. The tool matters less than the separation.) Each one got the same neutral question: what is the most material piece of information in these accounts?

All four came back with the same verdict. The cash is probably not real.

None of them could name the company.

And here is what happened in the real world, a year after that filing. The company’s audit firm tried to verify those bank balances. The company pointed the audit team to a suspected fake website for the bank. A bank advice supporting the interest income contained mathematical errors that management waved off as the bank’s clerical mistakes. The auditors resigned; I read the resignation letter which is on EDGAR, more on that later. The company lost its NASDAQ listing, and the SEC later revoked its registration.

Hold the name for now. It matters later that you have probably never heard of it.

Here is the problem, and it is probably yours too. Finding red flags was never the hard part. Any screener will hand you red flags all day: debt too high, margins too thin, receivables aging. But the market prices those within minutes of a filing hitting the wire, so a list of visible flags is worth almost nothing. The flags that destroy capital are the buried kind, the ones that only show up when two numbers from two different pages are forced to agree with each other and even when you actually found one, the question which nobody is going to teach you to answer is going to decide everything that is going to be the answer: is this flag survivable, or is it fatal?

A few weeks ago, a paid subscriber sent me an article about exactly this, with four questions attached. The article was by , who writes Al Smallwood. The questions were sharper than most professional research briefs I have received:

1. What do you think of this research method?

2. Would AI only find fraud after the fact, or can it uncover it before the news breaks?

3. Can you remember an accounting scandal from your analyst days, and show how AI would have found it right away?

4. Is this workflow worthy of the newsletter?

I told him that I would get back to him in a few days after thorough research. What follows is what I found.

I built this workflow and then ran it on three resolved cases. Out of that, one case was recognised by the model from the memory, second one they reasoned out cold, and in the last one, the workflow caught my own data errors before it actually caught the fraud.

The second question is the one everyone actually cares about, so here is the short answer. AI will not predict the next scandal, but if you run it correctly, it does something more useful. It will tell you how to differentiate a survivable problem from a fatal one.

There are two kinds of red flag: survivable and fatal.

The article my subscriber sent was built around a small UK online-dating company called Cupid plc. On paper, 2011 was a great year: revenue up 109 percent to £53.6 million, profit doubled, first dividend. The material fact was buried in a payroll footnote: 177 of the company’s 384 employees were classed as marketing staff, at a company whose advertising was almost entirely online. You do not need 177 people to buy online ads in a company of this size. In early 2013 the BBC, and a newspaper in Ukraine where most of the staff actually sat, reported what they appeared to be doing: posing as interested members to draw real users into paying to reply. The company denied any organised practice, and a review it commissioned from KPMG did not confirm one. I still think the footnote was telling the truth. The market did not wait for a verdict. On the allegations the shares lost more than half their value in a single day, and although they bounced when the company pushed back, the dating business never recovered. Cupid sold off the sites at the centre of it and was out of online dating within two years.

Now, notice what kind of flag this is. It did not say a particular number needed checking. Instead, it questioned whether the product itself was real. That split is the one thing no screener will ever be able to hand you. Every red flag belongs to one of the two species.

The survivable species says: a number needs to improve, or be verified. High debt. Weak cash conversion. Ugly margins. A one-time liability nobody can size yet. If the business underneath is real and the management honest, these flags resolve, and the pessimism they created is where the money gets made.

The first example I can think of is from 1963, when a subsidiary of American Express had issued some warehouse receipts for vegetable oil that did not actually exist. The tanks that supposedly held oil were actually holding seawater with a film of oil floating on the top. The actual claimed oil at that one facility exceeded the entire soybean and cottonseed oil supply of the entire country. When this got out, stock fell from $60 a share to about $35, but the actual receipts were only a one-time liability sitting on a balance sheet of an honest(debatable) functioning traveller’s check and charge card business. That damage had a lower floor. Warren Buffett put about 40% of his partnership into the stock in 1964 and roughly doubled his money when the actual claims were settled.

Now the same lesson from another market. In 2002, Titan, an Indian company, was worth about forty-five million dollars and carried debt of roughly three times its equity, dragged down by a weak watch business and a failed expansion into Europe. Profit fell by half in a single year. If you look at the numbers, it looked as if the company is already finished, but the debt was a bounded problem. It was not a lie. Underneath that debt was an honest, capable management and a jewellery brand Tanishq, which was already a top five retailer in the country and was growing fast. That is the survivable flag: a stretched balance sheet next to a real, growing business. Rakesh Jhunjhunwala met the incoming managing director, came away convinced the man was blunt and honest, and bought. Titan’s market value went from about forty-five million dollars to three and a half billion in a decade, and by the time Jhunjhunwala died in 2022 his stake was worth more than a billion dollars, roughly a thousand times what he had paid. Both stories are history, not recommendations. But they define the species: a bounded problem, honest people, a real business underneath, and the flag healed.

The fatal species says: the revenue is manufactured, the product is fake, management is lying. Cupid and the three names coming later in this piece. These flags do not resolve. They detonate. The uncomfortable part is that the fatal species usually wears the better-looking numbers, because the clean numbers are the lie.

Which is why the deciding variable is not in the ratios at all. Honest management with ugly numbers is usually survivable. Conflicted, unchecked management with suspiciously beautiful numbers is usually fatal. The honesty of the people running the company is the thing that decides it. That one judgment sorts every other red flag into survivable or fatal.

One honest caution before we go further. This workflow filters out the zeros. It does not find the Titans. Keep that in your head through everything that follows.

The hard part is choosing which flag is fatal.

So how do you build something that makes that call, survivable or fatal, on a filing full of flags? The problem is not the one you would expect.

Point a capable model at a set of accounts and ask it the open question, what is the most material fact here, and it finds plenty. It hands you weak cash conversion, a related-party balance, aged receivables. All real. All survivable. It rarely lands on the one buried thing that means the revenue is fake, even when that thing sits in the same document. The experiment that sparked this piece showed the pattern cleanly: split the reading across specialist agents and they find the fatal fact almost every time, but the manager that ranks their findings throws it away almost every time, because it sides with whatever the most agents agree on, and the fatal finding only ever shows up in one report.

Finding flags is easy. Choosing the fatal one out of a pile of survivable ones is the whole job, and it is the exact thing a human analyst is paid for.

So I built the workflow to protect that one minority finding at every step, instead of letting a vote bury it. Here is the whole thing, the way I ran it.

1. Feed it the filings, and nothing else. The annual report, the last few quarters, the prospectus if there is one. Not news, not summaries, not the company’s own investor deck. (If you have never wired filings into Claude Code, I walked through that setup in How I Set Up Claude Code as My Investment Research Analyst Start there.)

2. Split the work across five separate agents. Be literal about this, because it is the part that makes the whole thing work. You do not ask one model to do five jobs. You run five separate agents, each one a clean instance with a single task, and none of them able to see what the others found. In Claude Code these are five separate subagents, each in its own context, none of them able to see the others. Claude Cowork can split the job into parallel agents for you the same way. One agent reads whether the sales are real. One checks whether the per-customer and per-store numbers make sense. One asks whether the reported profit is backed by cash actually coming in. One hunts for two disclosed numbers that cannot both be true. And one looks at who controls the company and whether they can be trusted. Hand all five jobs to a single model and you just get its one worldview, five times over. Five points of view have to be built in on purpose. You cannot get them by asking one model to be broad.

3. Make each agent answer the same four questions. What is the single most important thing you found? What is the math behind it? Is it survivable or fatal, and why? And what one piece of outside information would prove it true or false? Each agent comes back with one finding, not a list, because a list is where you will see the fatal one hiding among all the safe ones.

4. Send each finding to a fresh agent to argue in full. This is the step that was missing everywhere else, and the one that makes the whole thing work. A new agent takes one finding on its own and builds the strongest case for it: assume this single thing is the key to the entire company, argue it all the way out, then give the ordinary explanation too. The point is to see each finding at full strength before it is ranked against any other.

5. Let one final agent judge by impact, not by how many agents agreed. Only now does a last agent, the judge, read the finished arguments and rank them by a single question: if this finding is true, how much does it actually hurt the company? Go back to Cupid. One finding sounds like a footnote: 177 of the staff are in marketing, at a company that advertises almost entirely online. The other sounds like a scandal: 2.3 million pounds of customer money is parked in a separate company the directors own. But the scandal only means one balance needs checking, while the footnote questions whether the revenue itself is real. Judged by what each would really do to the company, the footnote wins.

6. Every number reconciles to the filing, or it does not get used. This one sounds too obvious to state. It is the step that saved me from myself, and I will show you exactly how in the Luckin run.

Here is the instruction you give Claude, and it pays to be literal about it. The prompt names all five jobs and tells Claude to run each one as a separate agent, blind to the others. That is how Claude knows to use five agents, one per job, and not one agent doing all five, or two, or eight. You do not paste it five times. The only rule that cannot bend is that they stay separate. The moment all five share a single chat, they collapse back into one worldview, which is the exact failure this design exists to prevent.

Read only the filings below. Reason only from these numbers.
[the company's filings]
First, run five separate agents, each a forensic specialist, blind to
each other, one for each job:
1. Are the sales real?
2. Do the per-customer and per-store numbers make sense?
3. Is the reported profit backed by cash coming in?
4. Do any two disclosed numbers contradict each other?
5. Can the people who control the company be trusted?

Each of the five returns exactly:
- THE SINGLE MOST IMPORTANT THING YOU FOUND
- THE MATH BEHIND IT
- SURVIVABLE OR FATAL, AND WHY
- THE ONE OUTSIDE FACT THAT WOULD PROVE IT TRUE OR FALSE

Then take each of the five findings and argue it at full strength.

Finally, after all five are done, one more agent, the judge, reads the
arguments and ranks them by impact: if a finding is true, how much
would it hurt the company.

That is the whole machine.

Two notes on using it, because they answer the obvious question. You run this on a company you are actually thinking about owning, feeding the agents its real, current filings. That is the point of building it. Everything from here on, though, is me proving it works on three companies whose stories already ended, because a finished fraud is the only place you can check the method against an answer you already know. And on those three I added one step you would never bother with on a live company. I stripped the name and the country out of the numbers first, so I could see whether the agents were working the problem out or just remembering a famous scandal. On a company nobody has exposed yet, there is nothing to remember. You just hand it the filings.

The very first company answered that memory question before I finished asking it.

WorldCom was the perfect first test, until the models recognized it.

Start with the cleanest fraud in the record, because you can see the trick the moment someone points at it. Through 2001 and into 2002, WorldCom took an ordinary running cost, the “line costs” it paid other phone carriers to carry its calls, and recorded billions of it as though the money had been spent building the network instead.

Here is why that one move matters so much, in plain terms. A normal running cost comes straight off this year’s profit, all at once. But money spent building something long-lived, a piece of network, does not; you spread that cost out over many years. So by relabeling this year’s phone bills as long-lived assets, WorldCom kept most of the cost off this year’s accounts. Profit appeared where a loss was actually sitting.

Here is the tell, in WorldCom’s own filed numbers. Line costs rise and fall with how many calls run across the network, so they should hold a fairly steady share of revenue. WorldCom reported them at 41.0 percent of revenue in 1999, 39.6 in 2000, and 41.9 in 2001. The dollar figure for 2001 came back to 14,739 million, the same to the last digit as 1999, as if it had been set by hand rather than run up by a real business. And while revenue was flat to falling, the company kept pouring money into new equipment faster than its existing equipment was wearing out, so the asset side of the books swelled while the business shrank. The costs it was quietly not counting had to go somewhere. They went into that equipment line.

The company later admitted the size of it: 3.055 billion dollars moved out of line costs in 2001, and another 797 million in the first quarter of 2002. Add the 2001 piece back where it belonged and line costs were not the reported 41.9 percent of revenue but closer to 51, and the year’s profit becomes a loss. Later restatements pushed the total past nine billion dollars. No short seller found this. An internal auditor did, by reading the capital-spending accounts.

I gave four of the agents these real figures, with the name and the years stripped off, and one neutral question. The fifth agent, governance, reads the people who run a company rather than its accounts, so I held it back for the case where the numbers alone would not be enough.

All four came back fatal. Running costs are being moved onto the asset side of the books, the reported profit is not real, and the single document that would settle it is the list of exactly what the company counted as new equipment that year. Reading nothing but an anonymized income statement, the agents pointed straight at the one schedule where the real fraud was later found.

Then three of the four added a line I had not asked for. This is the WorldCom pattern.

My blind test was not blind. The models knew the answer the way you know the punchline of a joke you have heard before.

So I tried to beat the memory. I shrank every figure down to a fraction of its real size, so the company now looked like a mid-sized carrier rather than a giant, and I kept every ratio identical to the last decimal. Then I ran it again. Three of four named WorldCom anyway.

You cannot hide the shape of a fraud by changing the numbers. The model recognizes the setup even when every number is different.

A telecom hiding its call costs inside its equipment spending is WorldCom to any well-read model, the same way a payroll full of fake users is Cupid. I want to be fair about what that recognition is, though, because it cuts both ways.

In my test that recognition was a nuisance, because I was trying to prove the machine could reason, and a model that already knows the answer proves nothing. But on a real job it is useful. Run this on a company nobody has questioned yet, and one agent saying the cost pattern here looks like WorldCom is a genuine early warning. The model has effectively read every fraud that was ever written up, so it can spot a familiar setup that a busy analyst, seeing the filing for the first time, might read straight past.

It is still only a warning, not a verdict. The same reflex will also point at honest companies that simply spend a lot on equipment, so you still have to do the work of deciding which one you are actually holding.

The one agent that never said the name is the one that showed the real skill. It reasoned from the numbers alone and stayed careful. Probably fatal, it said, but a cost that behaves like this can sometimes have a dull and innocent explanation, so pull the equipment schedule and check before calling it. That agent was slower, and it was the most useful of the four. The three fast agents recognized WorldCom; the slow one actually did the analysis. That gap is why the workflow ends with a judge weighing the finished arguments, rather than counting how many agents agreed. It is the same discipline I wrote about in making the models reason from the documents, not their own memory.

Which is leaving the real question still standing. If you take the memory away, can it still do the work? For this exact question, I needed a fraud that the AI model could not have memorised.

Read more

I do not trade options anymore.

I used to. During COVID I traded them heavily, and if I am honest, my start was earlier and worse. Before my MBA I traded options without understanding how they worked. I sometimes made money. I understood nothing. Most of the options market still trades exactly like that, on greed and no understanding, and I do not think that will ever change.

The MBA taught me the machinery. Theta, delta, how these contracts actually price. I carried that straight into my job as an equity dealer on a broking desk for wealthy clients (the dealer is the person who actually advises and places the trades).

At my peak I was handling 450 client portfolios. Thirty to thirty-five of those belonged to serious derivative traders.

I worked on that desk for only two years, but the experience packed into them felt like ten or twenty. I had no personal life. Weekends went into learning, and the relationships that mattered to me paid for it. I am not recommending that trade. I am telling you about it because everything I actually know about options came from those clients.

They ran from 25 to 85 years old.

One was in his seventies, an entrepreneur with a business worth more than a billion dollars, and on the side one of the most successful option short sellers I have ever seen. Another was 27, a second generation entrepreneur, already trading very successfully. And there was a money market dealer, somewhere in his forties, who panicked every single day. Every morning he called four or five different dealers and analysts for opinions, and it did not matter whether the person had six months of experience or twenty years. His voice always sounded the same. Shubham, what do I do? What do you think? He was one of the most successful traders I served anyway. In between sat sellers who treated options as risky free money, hedgers, adrenaline traders, and people whose entire income came from the market. Many could not have explained what an option actually is. They just knew it made money.

I talked to these people every day, sometimes on weekends. Slowly I worked out that the successful ones shared the same few traits, even though they came from completely different fields and generations. It took me much longer to admit that those traits were missing in me.

My clients’ derivative books did well. My own did not. With client money there was a sense of responsibility. The risk stayed measured, and the difficult decision got taken immediately, without waiting. With my own money, emotions arrived, the risk grew, and the difficult decision waited. After enough losses, I accepted what that meant.

I stopped trading options. That was almost a year ago.

The trader in me is probably gone. The opportunistic gambler is still alive. Two or three times a year I take a leveraged futures position on a company I have very high fundamental conviction in, using my long term holdings, pledged, as the margin. The rest of the time I wait. Two months, four months, sometimes more than six, with no position at all. The itch to take a position arrives every week, sometimes every day. Controlling that itch, ignoring it, and then continuing to ignore it until an opportunity arrives on one of my watchlist stocks that is genuinely too big to ignore, is the single most difficult thing I do as an investor. The waiting is not something I have to force anymore. It is something I just do.

What remained under all of this was what I have always been, a fundamental investor. When I study a business I look at three things: what goes in, the revenue model, and what comes out. The inputs a company buys, the way it turns them into money, and the products or services that leave the door. That is the basic case study of any business. And for years I have believed that only two things decide long term returns, valuation and earnings growth. I still mostly believe that.

That belief gave me a comfortable reason to ignore the options market on the stocks I own. Whatever kind of week the options are having on a company, the price will eventually follow the earnings. So why look.

Options are not only for trading.

The positioning on a stock you hold carries information about that stock, and some of the people paying for that positioning know things you do not. Their positioning is public. Volume (how much traded today) and open interest (how many positions are still standing) print on every contract, every day, where anyone can read them. For most of my investing life, checking that daily was a full time job that belonged to someone else. In 2026, AI can do it in minutes a day. Ignoring it because long term investing does not need options felt reasonable for years. With AI in the picture, it is closer to being ignorant, and the cost is real alpha left on the table. Half a percent here, one percent there, sitting in things a machine can monitor while I read filings.

So, almost a year after I stopped trading options, I opened Claude Code and started building an options screener. What I wanted from it had nothing to do with trading.

A small warning before we go on. This edition is a little more technical than my usual ones. Every term gets explained in plain words as it appears, and if you have never touched an option in your life, that is fine. I wrote this for you too.

What the options market can tell you about a stock you already own

I call the tool the disagreement detector, and its job is one question: is the options market disagreeing with me about a business I own?

Every options tool I have come across is built for trading. Scanners that hunt the whole market for unusual activity, alert services that push the day’s biggest prints (a print = a trade that showed up in the data), all of it built to answer what should I trade today. I have no use for that question. I hold my companies for years, and I have no intention of trading a single option on any of them.

But I do care when someone starts paying serious money for a view that contradicts mine. A stock position is mostly silent. You cannot see who is accumulating shares against you until a 13F (the quarterly holdings report large fund managers must file) shows up, forty five days after the quarter ends, and even then only for the largest managers. An option is louder. Every option names the price level the bet turns on (the strike), the date the bet expires, and the real money paid up front for it (the premium), and the exchange publishes the volume and open interest on every contract, every day, for free. When positions like that appear on a company I own, persist, and keep growing, I want to know about it.

So the disagreement detector does not scan the market. It watches sixteen names: the businesses I own or cover, plus the kind of quality companies I would want to own one day. Every morning it takes a snapshot of the full options chain on each of them (the chain is the list of every option contract that trades on a stock) and writes me a short report. I call that output the disagreement report. Most days, for most names, the report says the same thing: quiet. Quiet is a finding. It means nobody put meaningful new money against my read that day.

And when the report is not quiet, the flag is a question, never an instruction. The detector surfaces what changed on which name and what to investigate. It does not tell me what to do about it, and nothing in this edition tells you what to do about anything either.

One honest caution before we go further. Most unusual options activity is not information. It is hedging, rolling, expiry mechanics, or a volatility trade with no view on direction at all. A tool like this earns its keep not by finding flags but by killing them. Hold that, because the killing is where this build gets interesting.

First, what it is built from.

How to build an options screener with free data and Claude Code

The build had one rule: free data only.

The professional version of this problem runs on paid data feeds. I wanted to know how far free public data could go before I paid anyone for anything. I ended up with three sources, and one honest gap.

Cboe, the options exchange, publishes the full delayed chain for any US ticker as a single file a machine can read: every contract, its volume, its open interest, the last trade, the bid and the ask. About fifteen minutes behind the market, free, no account, no paywall.

SEC EDGAR publishes insider filings. When an executive or a director trades their own company’s stock, a Form 4 must be filed within two business days, and the form carries a detail most people skip: a checkbox showing whether the trade ran under a pre-scheduled plan. Scheduled selling is near-noise. A discretionary trade is a decision somebody made that week.

The House of Representatives publishes its members’ trades, up to forty five days after the trade, as PDF filings. This is the ugly corner of the pipeline, and it is exactly where Claude Code earns its keep: it reads the PDFs and pulls the transaction rows out. About one in ten is a scanned image a computer cannot read cleanly. The tool counts those and says so, instead of pretending they do not exist.

The honest gap is the Senate. Its disclosure site turns away automated requests, so version one is House only, and every daily report states that in its header.

Then came the constraint that shaped the whole design. Free end-of-day data has no trade tape. I cannot see whether a trade hit the bid or the ask, and I cannot see sweeps, the aggressive multi-exchange orders that flow services are built around. So this tool does not pretend to. It asks three slower questions instead.

1. Did volume show up today far above the open interest that existed yesterday, which means positions were opened, not managed?

2. Is open interest growing day after day on the same contract, which means someone keeps paying to hold the position?

3. And is the balance of puts to calls (puts are bets on a fall, calls are bets on a rise) moving against this name’s own history, not against some other stock’s?

Slower suits me. A day trader needs the sweep the second it prints. A decade-holder needs to know whether somebody is still paying for a position against his thesis a week after they opened it. End-of-day data answers the second question fine.

The scan behind those three questions is plain arithmetic, and I am publishing the exact rules alongside this edition, in a public folder of files anyone can open. The floors: volume at least 2.5 times the standing open interest, at least 250 contracts, and roughly $200,000 of premium at stake before a contract even becomes a candidate, with smaller floors on thin chains so a quiet name like Solstice stays visible, and open interest growth of at least a thousand contracts across consecutive days before anything is called a build.

Every candidate then walks an exclusion ladder before it can be called a signal:

• Expiring within three days: almost always rolls or expiry mechanics.

• Earnings within about ten days: presumed hedging (buying protection, not making a bet), not information.

• Matching call and put volume at the same expiry: a volatility trade, no direction in it.

• Deep in the money with volume matching open interest: mechanics, not opinion.

• The whole universe leaning one way on the same day: the market repricing, not news about my company.

The machinery runs itself. Every morning my machine snapshots all sixteen chains at the prior close, automatically, whether I am paying attention or not. The archive of those daily files is the dataset. None of it can be recreated later, which is why the snapshots started the same day as the idea. And here is the prompt, exactly as I run it, every morning:

run the tape

That is the entire interface. Behind those two words sits a saved set of written instructions and a few small programs that Claude Code runs in order: take the snapshot, run the scan, pull the fresh Form 4s and the congressional PDFs, walk the exclusion ladder, and write the day’s disagreement report with every number traced to the snapshot file it came from. I described the design to Claude Code in plain English across two sessions on July 23, and it wrote those programs itself.

What does a snapshot look like? Here is the first one on file, the July 22 close, exactly as it sits in my notes:

Nobody reads this raw. That is the point of the whole build.

None of this, so far, finds anything. It only produces candidates. What decides whether a candidate is real is a test that runs overnight, and that test is the best trick in the whole build.

The overnight test that separates real option positions from day trades

Free options data has one clean trick in it, and the whole tool stands on that trick.

Two numbers, both public. Volume counts how many contracts changed hands today. Open interest counts how many contracts people are still holding when the day ends. Think of volume as today’s activity, and open interest as the bets that survived the day. The exchange updates open interest overnight.

Here is why that matters. Say an option that barely trades on most days suddenly trades thousands of contracts. That looks dramatic. But the volume number alone cannot tell you which of two very different things just happened. Maybe somebody opened a big new bet and is holding it, with real money now riding on it. Or maybe traders bought and sold all day, and by the close nobody kept anything. Both look identical in the day’s volume. The difference shows up the next morning. If the buyers kept their contracts, open interest jumps. If open interest does not move, nobody went home with a position, and the drama meant nothing. So the tool never trusts a dramatic day on its own. It writes the candidate down, and the next day’s run checks it against the open interest that settled overnight.

One practical note on the rhythm. I run the tool once a day, after the US market closes. The data capture is separate and automatic: every morning my machine stores the snapshot whether I show up or not. So there is no second run and nothing waits on me. Today’s run scores today’s new candidates, and it also opens with a checklist, judging yesterday’s candidates against the open interest that arrived overnight. Every report literally ends by writing tomorrow’s checklist for itself.

The very first run showed me how well this works. On July 23 the scan flagged eight contracts across five of the sixteen stocks. Each one had far more volume in a single day than all the contracts that existed before it, which is what new bets being opened looks like. One night later, every single one showed up in open interest. The buyers had kept them. I call that conversion: the day’s volume converted into bets somebody is still holding. Four of the eight, straight from that morning’s report:

Even the smallest of the eight, a bet against Intuitive Surgical expiring in mid 2027, kept about 200 of its 299 traded contracts. Nothing flagged that day was a day trade. Real people had opened these bets and were paying to keep them.

Notice what the test proves and what it does not. Conversion proves a bet is real. It does not prove the bet is smart, informed, or right. A real bet can still be simple protection, or a mistake. The test answers only the first question: is somebody actually holding this view with real money, or was it noise?

When a stock’s bets pass that test, the tool scores the stock. I call it the divergence score, and it is nothing fancy. It runs from zero to ten and just adds up evidence. The same bet growing for three days in a row earns points. Large money earns points. Several bets in the same direction landing on one stock in a single day earn points. An insider of the company trading the same way earns points too, but only when it was their own decision, not an automatic selling plan. A newly disclosed trade by a member of Congress earns one point at most, and only as a look backward: those filings arrive weeks after the trade actually happened, so they can quietly confirm a picture, never lead it. Then the score loses points for the things that usually explain away the drama: earnings coming up, a bet about to expire, or similar money sitting on both sides of the same stock. Five or more means the options market genuinely seems to disagree with me about that stock. Three or four means keep watching. Less than that, it just gets written down.

One caution I repeat in every report: the scoring is a hypothesis, not a proven system. Eight trading days of data is a start, not a track record. The tool writes down every scan, including the boring ones, exactly so the scoring can be tested against reality later instead of being trusted on faith.

The overnight test, though, needs no faith. A bet either stands the night or it does not. By the end of week two the test had killed five flagged candidates, including the biggest one-direction bet the tool had seen in the whole project: $17.9 million of buying on one stock, gone by morning.

That story is next.

Why unusual options activity is usually not a signal

On Tuesday July 28 the tool flagged the most dramatic thing it had seen. Somebody traded $17.9 million worth of bets that TSMC, the Taiwanese chipmaker, was about to rise.

The details made it look like serious money moving on serious information. The bets were call options: contracts that pay if TSMC climbs above 450 by mid September. 17,880 of them traded in one day, on a contract where only about 3,400 existed the day before. The stock had fallen 6.9% in a week, so someone appeared to be betting big on a bounce. Earnings were nowhere in sight, TSMC had already reported on July 16, so this was not the routine protection you see before results. And a second large bet sat right beside it: $3.25 million more, on a climb above 530. If you subscribe to any options alert service, this is exactly the kind of thing that arrives as a push notification implying that smart money knows something.

The disagreement report logged it, scored it 4 out of 10, and put one line on the next morning’s checklist: check whether the TSMC buyer kept the position.

The next day’s run opened that checklist. The count of contracts people were still holding: unchanged. 3,444 before, 3,444 after, even though 17,880 had traded in between. The second bet: unchanged too. Whoever traded roughly $21 million of TSMC bets that day bought and sold everything within the same day and kept nothing. The most dramatic bet of the whole project produced zero standing positions. The next day TSMC fell 4.5%.

Two days later came the punchline. TSMC jumped 7.6%. If that dead bet had been real, it would have paid handsomely. It was not there to collect. I wrote that down too, because the ledger records what happened, not what makes the tool look good.

By the end of week two the overnight test had killed five candidates:

The Accenture row deserves its own story, because Accenture is a stock I actually own, and because this kill humbled me twice in one week.

I researched Accenture earlier this year, so when the tool caught something enormous on it on July 23, I read that report more carefully than the others. What it caught: about 22,000 contracts betting on a rise and another 22,000 betting on a fall, landing the same day, expiring the same month. For scale, all the standing Accenture option contracts in existence totaled around 207,000, so one day of activity matched a fifth of everything that existed. If you turned those contracts into stock, they would control roughly $300 million worth of Accenture shares. Somebody very large had done something deliberate on my company, and I could not tell what.

The next day Accenture jumped 5.95%.

Be honest, you would have connected those dots too. Mystery trade lands Thursday, stock explodes Friday, the options market knew. I wanted to write that sentence. It is a flattering story. The report refused to let me, for a boring reason: end-of-day data cannot show who was buying and who was selling. The exact same trades could be a bet that Accenture would rise, or someone protecting a large holding against a fall. Both leave identical marks in the data. Unknowable, said the report. Check Monday’s open interest.

Monday’s answer: nothing. Both sides unchanged. Some forty four thousand contracts had traded and left no position behind, which means old positions were closed out or traders went in and out the same day. Not new money. The most cinematic trade of the project was dead, killed by the tool’s own test.

And then Accenture kept going anyway. Up 11% in two sessions. Up 18.7% in three. Up 24.8% in four. The reasons turned out to be plain fundamentals: a bigger share buyback, a dividend, AI related contract wins, and the whole IT services sector rallying together. None of that was ever in the options data. The tool had nothing to do with the rally and said so in writing. That is the second humbling, and the more useful one: sometimes the reason a stock moves was never in the options data at all. A screener that never kills its own ideas is just a tip sheet. This one killed the two most dramatic things it saw inside a single week, and that discipline is the only reason to believe it when it does flag something.

Because it did flag things. In the same two weeks, four flagged bets went on to pay, one of the tool’s cleanest signals expired worth nothing, and one disagreement is still open right now, against a stock that keeps rallying against it. The rest of this edition is that ledger, every number traced to its file.

Read more

I have taken losses in the market that stung for weeks.

The worst ones were the micro-cap bets. Extremely risky, and I knew it going in, and I took them anyway. When they broke, they broke hard. Sixty, seventy, eighty percent down, and I did what everyone does at those levels. I hoped.

It will come back. It always comes back. Look at the chart. One good quarter and I am whole again.

That voice is not analysis. It is anesthesia.

Some of those positions are still sitting in my portfolio today. I never sold them. I keep them there on purpose, red rows on the screen, reminders of the one lesson that took me years and real money to learn: taking a loss is not weakness. Accepting the fate of a position and booking the loss is the most courageous thing you can do in this business.

Those losses stung for three weeks. Four. Five. A couple of months, some of them. Then they faded. Today I could not tell you the exact numbers without opening my old statements.

This week I learned what a loss that does not fade feels like.

My cat Foxy died. She was two and a half years old. She showed me sides of myself I did not know existed, the good ones and the bad ones, things no market cycle ever surfaced in me. I will never forget her, and I do not think this one will ever fade. I am not going to compare her to stocks. There is no comparison. Nothing I have ever lost in the market comes close, and I want that sentence to stand on its own.

But loss clarifies. Sitting with it this week, I noticed that the regrets after any loss are always the same list. Maybe I should have done more. Maybe I should have been more careful. Maybe I should have never taken this risk. The list never changes. What changes is what you do while a loss is still open.

In the market, while a loss is open, you have exactly two options. You can hope. Or you can go read the documents and let them tell you what you actually own.

The market is running the hoping loop on one stock right now, at full volume.

On July 9, S&P cut Oracle to BBB-. One notch above junk. Oracle. The boring database company that sits in retirement accounts because nothing ever happens to it. The stock is down more than 60 percent from its peak, and everyone still holding it is running the loop I know by heart. It will come back. It always comes back.

Maybe it will. But hope is not a process.

Oracle filed its annual report on June 22, seventeen days before the downgrade. It contains $638 billion of contracted future revenue. A year ago that number was $138 billion. Half a trillion dollars of new promises in twelve months.

So I did the only thing I trust when everyone is shouting. I pointed Claude Code at the fresh filing and had it search every page, every exhibit, for the name of the customer behind roughly half of that number.

Zero results.

Everyone Is Arguing About Oracle. Almost Nobody Has Opened the Filing.

Here is what this week sounded like if you hold Oracle.

Start with the downgrade itself. S&P’s reasons for the cut, in plain words: Oracle plans to spend $90 to $95 billion on data centers this fiscal year, its free cash flow is projected at negative $42 billion, and roughly half of its giant backlog depends on a single customer. One detail most of the takes skipped: the outlook S&P attached to the new rating is stable, not negative. Read both halves of that. The agency is worried enough to cut, and not worried enough to signal the next cut.

The stock made a new 52-week low at $123.66 on Thursday. From its September peak, it has lost more than 60 percent.

And the market cannot agree on what any of it means. The same week the downgrade landed, one Wall Street firm published a note arguing the stock triples from here. Bears are calling Oracle the first domino of the AI buildout. Both sides post charts. Both sides sound certain.

They are not just posting, either. They are betting. On Polymarket there is a live market on the question underneath every Oracle take, whether the AI bubble bursts, and real money moves its odds all day.

Watch the odds tick for a minute. Then notice what they are made of. Every tick is opinion, priced in real time. Not one of them is a new fact about Oracle. The facts are sitting where they were filed, largely unread.

Scroll through the takes and you will notice something.

Almost every argument, bull or bear, is built from the same three ingredients. The downgrade headline. The stock chart. And a number Oracle announced on an earnings call. Almost nobody quotes the document underneath all of it.

Oracle filed its FY2026 annual report on June 22. Seventeen days before the downgrade. It is the most information-dense thing the company has published all year, and it answers on the record, under liability, with numbers someone signed.

Not every question. Filings never answer every question. And this one refuses to answer the biggest question you can ask of it: who, by name, is behind the $638 billion? Hold that question. It is the entire second half of this piece.

So that is where I went. Not to form an opinion first. To build a map.

Who supplies Oracle. Who buys from Oracle. How much hangs on whom. The questions you would ask about any business whose price just got cut by more than half.

It turns out there are exactly five of those questions. And they work on any company with a filing.

Five Questions. The Company Is Just the Parameter.

Here is the whole framework in one breath. Five questions, in a fixed order, and the order is the discipline. Each question becomes one prompt that forbids the machine from guessing. The company is just the parameter.

1. Who does it depend on to build? Dependence hides on the balance sheet, not in the supplier list.

2. Who does it depend on to get paid? Future revenue is a promise, and someone specific made it.

3. What is the receipt behind every name? Disclosed, on the record, reported, or NOT DISCLOSED. Never an estimate dressed as data.

4. Who appears on both sides? When your supplier is also your customer, the demand signal needs a second look.

5. What can the map not show? Last on purpose. The silences are the risk list.

Build side first, because that is where disclosure is forced. Paid side second, because that is where it thins. Receipts before meaning. Both sides only after both maps exist. And the gaps last, always last, because the gaps are the finding.

What the five questions produce is a picture professionals will recognize. Bloomberg terminals call it SPLC: the company in the center, suppliers on one side, customers on the other, the dependencies quantified. That shape is the right shape. The framework is not mine and it is not new. What is new is that you can build it yourself from the primary record, and hold every box on it to a receipt.

If you have never set Claude Code up for this kind of work, the whole install is in How I Set Up Claude Code as My Investment Research Analyst. Start there. One caution that never changes: give it a folder with the filing inside, not your whole vault, and let it work in that sandbox.

Now the run. Oracle, the loudest stock in the market, fresh 10-K on the table.

STEP ONE: WHO DOES IT DEPEND ON TO BUILD?

The principle: dependence hides on the balance sheet. A supplier list tells you who a company shops with today. The commitments, the leases it has signed but not started paying, the power it has promised to buy, the borrowings of other companies it has guaranteed, tell you who it is chained to for the next twenty years.

The prompt I gave the tool:

read Oracle's FY2026 10-K in this folder. I want the build side of its dependency map. from the filing only: every supplier dependence it discloses, chips, manufacturing, data center capacity, power, and the commitments that lock those dependencies in: leases signed but not yet commenced, unconditional purchase obligations, guarantees of anyone else's borrowings, and the financing raised for the buildout. exact quote and section for every item, accession number on the file. if a counterparty is not named, write NOT DISCLOSED, do not guess who it is. save it as a markdown file, then re-read every cited section from scratch and flag anything that does not reconcile.

What it returned:

The build side is enormous, and almost nobody in it has a name.

Chips: the filing discloses dependence on "AI accelerators, including graphics
processing units" and admits Oracle "at times [has] to accept less favorable
terms with suppliers to minimize supply constraints." Supplier names: NOT
DISCLOSED. Zero occurrences of any chipmaker's name in the document.

Data centers: $260 billion of lease commitments that are signed but have not
even commenced, expected to start between fiscal 2027 and 2029, with terms of
fifteen to nineteen years. That sits on top of $30.2 billion of lease
liabilities already on the balance sheet, up from $13.4 billion a year ago.
Landlord names: NOT DISCLOSED.

Power: $13.3 billion of unconditional purchase obligations, "primarily related
to data center power arrangements." Counterparties: NOT DISCLOSED.

Signed after the year closed: another $19 billion of purchase commitments for
cloud infrastructure, five-year term. And one guarantee: Oracle has guaranteed
up to $3.3 billion of a data center landlord's borrowing, which matures in
September 2026. The landlord: NOT DISCLOSED.

The money behind it: capital expenditures went from $21.2 billion to $55.7
billion in one year, and the company issued $43 billion of senior notes in
fourteen tranches, with maturities running out to February 2066.

One sentence flagged from the risk factors, because it ties the whole side
together: the lease terms "typically do not align with the duration and pricing
of customer contracts."

Read that back slowly, because the shape of it is the finding. The filing is precise about the sizes. A quarter of a trillion dollars of leases that have not started. Forty-year money funding them. And it is silent, almost completely, about the names. On the build side of the map, the amounts are disclosed and the counterparties are not.

Hold that asymmetry. The paid side, next, repeats it at a scale that is hard to believe. And then, in a place nobody thinks to look, it flips.

STEP TWO: WHO DOES IT DEPEND ON TO GET PAID?

The principle: future revenue is a promise, and someone specific made it. A backlog is not money. It is a list of counterparties who have agreed to pay, and the only thing that makes the number worth anything is who is on the list. So the second question walks to the other side of the map and asks the filing to name the promise-makers.

This is also where the question I asked you to hold at the top of this piece stops being rhetorical.

The prompt I gave the tool:

same filing, other direction. who does Oracle depend on to get paid? from the 10-K only: total remaining performance obligations and the prior year number, how fast the book converts to revenue, every customer concentration disclosure including the 10 percent revenue test, any admission of being concentrated among large customers, and every customer name that appears anywhere in the document. exact quotes and sections, accession number on the file. where the filing is silent, write NOT DISCLOSED. save as markdown, then re-read every cited section from scratch and flag anything that does not reconcile.

What it returned:

The book: remaining performance obligations of $638 billion, against $138
billion a year earlier. Multiplied by 4.6 in twelve months. The filing's entire
explanation for half a trillion dollars of new promises is one clause: "certain
significant cloud contracts that were entered into during the period."

The speed: about 12 percent expected to become revenue in the next twelve
months, roughly $77 billion. Then 34 percent between months 13 and 36, and 34
percent between months 37 and 60. A year from now, nearly ninety percent of the
book will still be a promise. And the book is not even complete: the filing
elects a permitted exemption that leaves certain consumption-based amounts out
of the $638 billion entirely.

The tests: the 10 percent customer revenue test comes back clean, all three
years. And one risk factor later, the filing concedes concentration among a
number of large customers anyway. Both sentences pulled verbatim.

The split of the $638 billion by customer: NOT DISCLOSED. Nothing anywhere in
the document quantifies any counterparty's share.

The names: zero. A full-text pass of the document and every exhibit returns no
customer name at all. The names behind the biggest number in the filing do not appear in the document that printed it.

Here are those two sentences, and I want you to read them the way they sit in the document, a few pages apart, saying opposite things.

Oracle’s filing states, on the record: “No single customer accounted for 10% or more of our total revenues in fiscal 2026, 2025 or 2024.”

That sentence is true. And it is useless. Both at once, and seeing why is the lesson of this whole step, because it transfers to every company you will ever read.

The 10 percent customer test looks backward, at revenue. Oracle’s FY2026 revenue was $67.4 billion, and the test patrols that number, the money already recognized. But the risk is not sitting in last year’s revenue. It is sitting in the $638 billion of promises, a book nearly ten times the size of revenue, and in that book there is no naming rule at all. The only figure anyone has put on what hides there is S&P’s, from the downgrade note: roughly half of the book depends on a single customer. Not half of a quarter’s revenue. Half of the book of promised future revenue, resting on one counterparty, and the institution willing to write that down is a rating agency, not the company.

And the filing itself concedes the gap, one risk factor later: “In certain OCI offerings, we are more concentrated among a number of large customers, which could increase these risks.”

Read those two sentences together. The first is what the disclosure rule forces. The second is what the risk actually looks like.

The revenue test cannot see a risk that has not become revenue yet.

By the time any of these customers is large enough in recognized revenue for the rule to force a name into the filing, the dependence will already have been sitting in the leases, the debt, and the data centers for years. That is the disclosure gap. Once you see it, you cannot stop seeing it.

Step One gave us amounts without names. Step Two gives us the biggest amount of all, and not one name. But a company speaks in more places than its annual report, and what it says elsewhere turns out to be the strangest part of this whole map.

That is Step Three.

STEP THREE: WHAT IS THE RECEIPT BEHIND EVERY NAME?

The principle: an estimate dressed as data is worse than a gap. A gap tells you where the record ends. An estimate hides that the record ended. So before any name goes on the map, it gets a receipt, and the receipt has exactly four grades. DISCLOSED, in a filing, with the quote and the accession number. ON RECORD, said by the company anywhere it speaks: earnings releases, calls, its own press page, offering documents. REPORTED, said only by the press, with the outlet named. Or NOT DISCLOSED, and the map says so instead of guessing.

The annual report went quiet on names. But a company speaks in more places than its annual report, and every one of those places is searchable.

The prompt I gave the tool:

now build the receipts ledger for every dependency from the last two steps. for each counterparty and each direction: is the relationship, and its size, disclosed in a filing, on the record from the company anywhere, earnings releases, calls, its own press releases, offering documents, or only reported in the press, and by whom. search Oracle's entire SEC filing history in the full-text index, and search the spelling variants of every name, not just the clean spelling. one line per name per direction, the strongest source, dated. never estimate a number, never average press reports. if nothing supports a size, the size is NOT DISCLOSED. save the ledger as markdown.

What it returned:

Build side receipts:

AMD, 50,000 MI450 GPUs for Oracle's cloud: ON RECORD, Oracle's own press
release, October 2025. Solid line.

NVIDIA, a reported $40 billion order of GB200 chips: REPORTED, Financial Times,
May 2025. Never confirmed by either company. Dashed line.

The landlords behind $260 billion of leases, the power counterparties behind
$13.3 billion of obligations, the borrower behind the $3.3 billion guarantee:
NOT DISCLOSED, every one.

Paid side receipts:

The full-text search across every Oracle filing since 2001, spelling variants
included, returns exactly four documents that name the customers. None of the
four is an annual report.

Four documents in a quarter century of filings. I want to walk you through them in order, because the order tells its own story.

The first is from June 2024, an earnings release celebrating a fresh contract, “one with Open AI to train ChatGPT in the Oracle Cloud.”

There is the name. And notice how it almost got away: spelled with a space, invisible to a clean-spelling search. That is why the prompt forces the variants. Exact-match search lies by omission.

The second is from March 2025, another earnings release, and this one is the sentence the whole paid side of the map hangs on: “We have now signed cloud agreements with several world leading technology companies including: OpenAI, xAI, Meta, NVIDIA and AMD.” Five names, from Oracle itself, in a filed document. Amounts: none.

The third, September 2025, mentions OpenAI only in a product sentence, a list of chatbots its database can talk to.

And the fourth is the one I keep coming back to. It is not a report to shareholders at all. It is a marketing document from Oracle’s February bond sale, filed with the SEC, telling the people it wanted to borrow from: “Oracle is raising money in order to build additional capacity to meet the contracted demand from our largest Oracle Cloud Infrastructure customers, including AMD, Meta, NVIDIA, OpenAI, TikTok, xAI and others.”

Six names and an “and others,” when Oracle is selling bonds. One clause, “certain significant cloud contracts,” when it files the annual report it is liable for. And the earnings release for the $638 billion quarter itself, this June: zero names. The pattern is not subtle. The names appear when the company is celebrating or borrowing. They vanish from the documents that carry the most liability.

Which leaves the size. For the largest customer relationship on the map, exactly one number exists in public. The Wall Street Journal reported it in September 2025: $300 billion of computing power over roughly five years. Two weeks later, OpenAI put it in writing on its own blog: a partnership that “exceeds $300 billion” over the next five years. Oracle has never said that number. Not in a filing, not in a release, not on a call. The biggest figure in this entire story has been confirmed by exactly one of the two companies involved.

The customer.

So now the ledger is full, and the map draws itself from it. Every line style below is a receipt grade, nothing more.

Read the left side first. Two boxes carry filing numbers, and they are enormous: the $260 billion of leases that have not commenced, and the $13.3 billion of power. The chips above them split by receipt: AMD solid, NVIDIA dashed. Now the right side. Six names, and five of the six boxes say NOT DISCLOSED. That is not my map being lazy. That is the actual state of the public record on $638 billion of promised revenue. And the orange box holds the one number on the record, confirmed only by the company that owes it.

One more thing before we move on. Look at which two names appear on both sides of the map. NVIDIA and AMD sell to Oracle on the left and buy from Oracle on the right. Hold that. It is the entire next step.

STEP FOUR: WHO APPEARS ON BOTH SIDES?

The principle: when your supplier is also your customer, the demand signal needs a second look. Money flowing in a straight line is easy to read. Money flowing in a circle is not, because some part of the boom on each side of the arrow is financing the boom on the other. This step exists to find the circles, and no supply chain tool I know of asks for them.

I told you to hold one observation from the map. Time to pay it off.

The prompt I gave the tool:

put the two sides of the ledger next to each other. which names appear on both sides? for each one: what flows out, what flows in, and the receipt grade on every arrow. then pull everything in the filing about financing between Oracle and its counterparties, in both directions: customers prepaying Oracle, Oracle extending payment plans to customers, Oracle guaranteeing anyone else's borrowings. exact quotes and citations. then line up the durations: the debt, the leases, the customer contracts. save it as markdown.

What it returned:

Two names sit on both sides of the map.

NVIDIA: money out for chips, a reported $40 billion order, dashed line. Money
in as a named cloud customer, size NOT DISCLOSED.

AMD: money out for 50,000 GPUs, on the record in Oracle's own release. Money in
as a named cloud customer, size NOT DISCLOSED.

The financing, in the filing's own words, one clause, cutting both ways:
certain customer contracts carry a significant financing component "either
because the customer has made significant prepayment before the corresponding
performance obligations are delivered or because we have provided long-term
payment plans to the customer."

Oracle has also guaranteed up to $3.3 billion of a data center landlord's
borrowing, maturing September 2026.

The durations: $43 billion of senior notes issued this year in fourteen
tranches, maturities to February 2066, coupons up to 6.85 percent. Total
borrowings up 40 percent in a year, $92.6 billion to $129.5 billion. Leases of
fifteen to nineteen years, most not yet commenced. And the one customer
contract with a public size: reported at roughly five years.

So Oracle pays NVIDIA for chips, and NVIDIA pays Oracle for cloud. Oracle pays AMD for chips, and AMD pays Oracle for cloud. Every piece of that is sourced, and nothing about it is improper. But be honest about what it does to your ability to read the business from outside. The demand on the right side of the map is partly coming from the companies being paid on the left side of it. That is not an accusation. It is a reading difficulty, and no rule requires anyone to resolve it for you.

Now the clause, because I want you to see both of its arms in the open.

Prepayment is the good direction. The customer pays before the work is done, the cash sits in Oracle’s hands, and the customer carries the risk. And on the June earnings call, Oracle put a number on that direction: roughly $75 billion of the book is prepaid or customer-supplied hardware. Real cash, real chips, already delivered.

The payment plans are the other direction. Oracle lending its own customers the time to pay for what they promised. Which contracts, and how large, the filing does not say.

That second arm has a history, and the history is why I refuse to shrug at it.

In the late 1990s, the companies selling telecom equipment started financing the customers who bought it. Lucent, the biggest equipment maker in America at the time, committed more than $8 billion in loans and guarantees so that young carriers could keep buying Lucent gear. The sales were real on paper. The backlog looked magnificent. Then the funding markets closed and the customers started dying. Winstar, a carrier that had signed to buy up to $2 billion of Lucent equipment with Lucent itself financing the purchase, went bankrupt in April 2001, and Lucent took an $845 million charge to fully reserve what that one customer owed it. The next fiscal year Lucent lost $16.2 billion and the stock fell below a dollar. The equipment makers had quietly become lenders to their own demand, and when the borrowers died, the backlog turned out to be a mirror, not a market.

I am not saying Oracle is Lucent. Oracle’s named customers include some of the most valuable companies on earth, a fair slice of its book is prepaid rather than lent, and the differences matter as much as the rhyme. The rhyme is not the size and it is not the outcome. The rhyme is the structure: a supplier whose reported demand depends, in part, on customers it also finances.

Now line up the durations the ledger pulled, because this is where the circle stops being a curiosity and becomes the risk. The debt runs to 2066. The leases run fifteen to nineteen years, and most have not started. The one customer contract with a public size is reported at about five years, and Step One already flagged the filing’s own admission that the lease terms typically do not align with the customer contracts sitting on top of them.

Forty-year money, funding twenty-year leases, resting on five-year promises.

Every arrow in that sentence carries a receipt. What none of the receipts can tell you is whether the promises will still be there when the leases commence. That question belongs to the one box on the map no filing can reach.

That is the last step.

STEP FIVE: WHAT CAN THE MAP NOT SHOW?

The principle, and the reason this step comes last: a filing’s silences are information too. Four steps built a map out of receipts. The fifth step reads the holes in it, because on this map, the holes are where the risk lives. Most research hides its gaps. This workflow ends by listing them.

The prompt I gave the tool:

last step. from the saved files only: everything this map cannot tell us. every material NOT DISCLOSED, every private counterparty with no public financials, every term the filing references but does not contain. then a watch list of dated, checkable items that would change the map, tied to the next filings. no opinions, no recommendations. save it as watch-list.md.

What it returned:

The map cannot tell you how the $638 billion splits by customer. No document
discloses it. S&P's "roughly half" is a rating agency's estimate, the closest
thing to a number that exists.

It cannot tell you the contract terms. Termination rights, take-or-pay
clauses, remedies if a customer walks, and which contracts carry which arm of
the financing clause. All NOT DISCLOSED.

It cannot tell you when the $260 billion of leases starts converting from
commitment to cost, beyond "fiscal 2027 to 2029."

And it cannot tell you anything, at all, about the financial condition of the
counterparty in the orange box. OpenAI is private. It files no 10-K. It
publishes no audited statements. One whole side of the biggest dependency on
this map sits outside the reach of every verification tool this workflow has.

Sit with that last one, because it is the finding under all the other findings.

Here is what the public record offers on that counterparty instead, every piece of it company-stated or reported, none of it audited: revenue of about $2 billion a month, roughly $24 billion a year. Not profitable. A $122 billion funding round closed in March at an $852 billion valuation.

Now my arithmetic, from those figures. A commitment that exceeds $300 billion over five years averages more than $60 billion a year, payable to one vendor, by a company currently generating about $24 billion a year in total revenue. That gap is the entire Oracle debate. And no document on either side of it lets you close the question, because one side of it does not file documents.

Oracle’s filing cannot name any of this. But read its risk factors knowing what you now know, and you can watch the lawyers describe it anyway. “Some of our customers may be highly leveraged and subject to their own operating and regulatory risks.” If key customers cannot perform, Oracle “could be locked into multi-year commitments for excess data center space and related capital expenditures, as well as associated financings, without receiving corresponding revenue.” And the quiet one, the accountant’s version of a scream: changes in key customers’ ability to pay could require “evaluation of the recoverability of related long-lived assets.” In plain English, if a big customer walks, Oracle tests its data centers for write-downs.

The market has already rehearsed this once. On April 28, a single report that OpenAI had missed user and revenue goals knocked Oracle’s shares and pushed the cost of insuring Oracle’s debt to two-week highs. One reported miss, by a company that discloses nothing, moved the price of Oracle’s risk. That is what it means to have half a backlog resting on a counterparty nobody can audit.

So the last output of the workflow is not a verdict. It is a watch list, saved as a file next to the map, of dated, checkable things that would change the picture: the backlog print each quarter and the 12 percent slice actually converting. Capex against S&P’s $90 to 95 billion projection, and free cash flow with it. S&P’s own published tripwires for the next cut, leverage sustained above 4.5x or no credible path to positive free cash flow by fiscal 2029. The $260 billion of leases as they commence. The securities class action filed in February about the cloud infrastructure business, because discovery has a way of surfacing what filings do not. And the simplest one: whether any future Oracle document ever attaches a name to the backlog, or whether the company in the orange box ever starts filing documents of its own.

The day either happens, half this map lights up.

There is a live market on exactly this. On Polymarket, real money trades all day on when OpenAI files to go public.

The day that market resolves yes, the dark side of this map starts filing documents, and half the boxes light up. Until then, the odds are the crowd’s best guess at when the lights come on. A guess is not a disclosure.

NOT DISCLOSED is not an empty cell on this map. It is the finding.

What the Map Told Me, and What It Could Not.

Step back from the five steps and look at what actually happened here.

The machine read a 10-K, then a quarter century of filings, then the public record around them, and built a map where every line is a receipt. It found customer names in a bond-sale document that the annual report withheld. It caught a name hiding behind a spelling. It put NOT DISCLOSED on nine boxes and never once filled a silence with an estimate. That part is real, it is repeatable on any ticker, and it is yours now.

But a map is not a verdict. And I owe you the other side of what the documents say, because the receipts cut in Oracle’s favor too.

This is not a company breaking on its operating numbers. The same filing that holds the $638 billion shows operating cash flow of $32 billion, up 54 percent in a year. Cloud infrastructure revenue grew 77 percent, to $18.1 billion, and cloud is now more than half of everything Oracle sells. The near slice of the book alone, the roughly 12 percent expected to convert within twelve months, is bigger than the entire company’s revenue last fiscal year. And the agency that cut the rating attached a stable outlook while doing it.

So hold both pictures at once, because both are true, and neither cancels the other. A cash machine growing at a speed no company this size has any right to. And a book of promises that is enormous, concentrated, unnamed, and partly unverifiable. The downgrade was never about what the business earns today. It is about the shape of the promises, and whether they will still be promises when a quarter trillion dollars of leases starts commencing.

What the machine could not do is decide any of that. It cannot tell me whether the company in the orange box performs for five years. It cannot tell me whether forty-year money against five-year promises is discipline or hubris. It cannot tell me whether to own the stock. That was never its job.

This is not a recommendation. It is a demonstration.

What I can tell you is what I trust. The map, because every line on it is a receipt. And the watch list, because every item on it is dated and checkable, so when the next filing lands, the machine reads it against both and the picture updates in minutes. That is the whole trade this edition offers: not my opinion about Oracle, but a way to make sure every opinion you hold can point at its receipts.

Trace It Yourself.

Everything in this piece lives in files, not memory. The filing quotes, the accession numbers, the receipts ledger, the watch list, the map. Every figure traces to a file, and every file traces to a primary source. If a number bothers you, trace it. The 10-K is free on EDGAR, accession 0001193125-26-277521, and every sentence I quoted is findable with your browser’s search box. That is the standard this newsletter holds itself to, and it is the standard I want you to hold me to.

And the five questions are yours now. Who does it depend on to build. Who does it depend on to get paid. What is the receipt behind every name. Who appears on both sides. What can the map not show. Point them at the company everyone in your feed is arguing about, in that order, and demand a receipt at every step. The map you get will be smaller than the ones the estimate-dressed-as-data crowd publishes.

It will also be true.

The thematic parent of this piece: Is Big Tech’s AI Capex Mispriced? I Checked, with Claude Code. That edition built the verification discipline this one runs on. If your own stock is the one selling off, the triage lives in How to Tell If Your Stock Is Weak or Your Entire Sector Is Shifting. And the NOT DISCLOSED discipline was born in I Taught Claude to Pull a Company’s Whole Debt Structure from One Filing.

The setup that started it: How I Set Up Claude Code as My Investment Research Analyst.
The memory layer that makes it compound: the Obsidian edition.

*Disclaimer: Alpha with AI covers US-listed and other global securities for educational and research purposes only. Everything in this edition is my personal view. None of it is investment, legal, or tax advice, a recommendation to buy, sell, or hold any security, or tailored to anyone’s situation, and I am not a registered investment adviser or research analyst. Oracle appears here as a demonstration of a research method, not as a view on what you should do with the stock. I may hold or take positions in securities discussed, and my views can change without notice. Every figure traces to a primary public filing or a named public source, accession numbers on file, but filings can be misread and reported figures can be wrong, so verify against the source before acting on anything. AI tools, including Claude Code, were used in the research and drafting of this edition, with every number checked back against the filed documents. Investing carries risk, including the loss of capital. This content is not directed at any person in any jurisdiction where its distribution would be contrary to local law. Do your own research and consult a qualified professional in your jurisdiction before making investment decisions.*

Written by Shubham Borkar | Research & Insights by Shikshan Nivesh AI Team

Financial Clarity. Insightful Ideas.

One more thing. The referral program is live. If you know one investor who was staring at a red screen this week, send them this edition. Referrals earn you months of paid access, and they are honestly the single biggest way this newsletter grows.

Refer a friend

On the banks of the Ohio River in southern Illinois sits a small town with a fifteen-foot bronze Superman in its square. The town is called Metropolis, and it has been officially calling itself his hometown since 1972.

Metropolis has a second claim almost nobody knows, and it is better than the first.

On the edge of town sits the only commercial plant in America that can take mined uranium and turn it into the gas every enrichment facility needs before it can become reactor fuel. Not the biggest such plant. The only one. Every American plan to build more nuclear power, every promise to feed AI’s appetite for electricity with reactors, every law weaning the fuel supply off Russia eventually runs through one industrial site in Superman’s town. The place famous for a hero who draws his power from the sun quietly runs on uranium. (No, they do not make kryptonite. I checked the product list.)

The town square celebrates a fictional man who could split atoms with his bare hands. The industrial park quietly feeds the machines that split real ones. Tourists photograph the statue every day. Nobody photographs the plant.

For decades nobody photographed the company either. The plant was buried inside Honeywell, one conglomerate asset among hundreds, filed somewhere between cockpit computers and thermostats. But conglomerates went out of fashion. Investors decided they would rather buy their businesses one at a time, priced one at a time, seen one at a time. So in October 2024 Honeywell started taking itself apart, announcing it would spin off its advanced materials business, and within months the whole empire was slated to split in three. The pitch, repeated in every breakup of the last decade: let each business finally be seen clearly.

Being seen clearly. Hold on to that phrase. It is the entire point of this story.

The chemicals business took the name of its own refrigerant brand, Solstice, and on October 30, 2025, Solstice Advanced Materials began trading on its own. And what walked out of that divorce is the most mild-mannered business you could invent. The gas inside your air conditioner. The film sealed over the pill you swallowed this morning. A featherweight fiber that stops bullets for a living and never once brings it up at parties. Clark Kent, incorporated in Delaware. Except this Clark Kent owns the plant in Superman’s town, and for eight months, almost nobody looked twice.

Then came the reveal. On June 4 this year, management held a webinar for one purpose: to show the world the nuclear business. Bank analysts dialed in with questions. The story started traveling. Therefore the stock climbed, and by the end of June it closed at $88.60, within a few percent of its high since the spin. The glasses were coming off, and the crowd was gasping on cue.

But thirty-two days after the reveal, on July 6, the same management announced a $14.5 billion agreement to acquire Element Solutions, an electronics chemicals company, paid for in cash it is borrowing and shares it is printing.

The crowd did not gasp. It ran. Down 15 percent on announcement day, by my math from the daily closes. Down 31 percent from that June close by Wednesday, $88.60 to $61.01, before clawing back to just under $63 as I write this. Type “why is SOLS stock down” into a search bar that week and every answer said some version of the same thing: dilution, debt, a company barely out of the wrapper betting itself on a mega-deal.

So here is the whole five weeks, drawn as the comic strip it secretly is. Panel one: the mild-mannered chemicals company steps out of the phone booth, cape visible, crowd gasps. Panel two: the crowd falls in love. $88.60. Panel three: the hero announces he is spending $14.5 billion, mostly borrowed, on a second costume. Panel four: tomatoes.

Panel five is still blank. Either Metropolis is watching its superman get booed for doing something heroic, or it just watched a man in a costume promise to fly.

The CEO went on CNBC the evening of the announcement, called it a generational growth opportunity in semiconductors and advanced electronics, and blamed the selling on hedge funds running merger arbitrage (funds that trade the gap between deal terms and market prices, not the business itself). The sellers did not go on television. Sellers never do. Panel five does not get drawn on television at all. It gets drawn in the filings.

So I pointed Claude Code at EDGAR and had it pull the entire paper trail of an eight-month-old public company: the merger agreement, the spin-off information statement, the tax agreement Honeywell still holds over its former child, both companies’ annual reports, every 8-K since the separation.

(The rig doing this work, Claude Code running inside Obsidian with every filing it pulls saved into the vault, has its own editions: how I set up Claude Code as my investment research analyst, and the 2.0 rebuild that moved everything inside Obsidian. Start there if you want to run this yourself.) And I gave it one instruction before anything else: build the case against this company before you let me like it.

What came back includes a number I have not seen in a single headline about this deal. It sits in the termination fees of the merger agreement, it is $513 million, and it explains why this deal is shaped the way it is shaped, why the terms could not bend while the stock fell, and why Honeywell, eight months after the divorce, is still in the room.

Hold that. It gets its own section.

And hold the plant in Superman’s town. Because the strangest part of this story is not the deal everyone is arguing about. It is the asset almost nobody was pricing before the argument started.

First, the company itself. I promise you touched its products this week without knowing it.

That, after all, is what mild-mannered means.

THE 60-SECOND THESIS

What they do. Solstice sells $3.9 billion a year of chemistry most people touch daily and never see: refrigerant gases for air conditioners (the largest line, $389 million in the first quarter alone), high-barrier film for pharmaceutical packaging, bullet-stopping fiber for armor, ultra-pure materials for semiconductor fabs, and uranium conversion for the nuclear fuel cycle, through the only plant in America that can do it. Two reporting segments: Refrigerants & Applied Solutions, the bigger one, earning 35 percent EBITDA margins last year and 39 the year before, and Electronic & Specialty Materials at about 18. The whole company earned $957 million of adjusted EBITDA last year at a 24.6 percent margin, and until July 6 it carried a clean balance sheet: about 1.4 times net leverage (net debt divided by yearly EBITDA, meaning how many years of profit it would take to repay the debt).

Why it is down. On July 6, management agreed to buy Element Solutions, an electronics chemicals company, for $14.5 billion including its debt. To see what the market flinched at, look at how the bill gets paid. Each Element share receives $10.00 in cash plus half a Solstice share. The cash half is borrowed: a $4.7 billion bridge loan from Goldman Sachs, which takes Solstice from owing about 1.4 years of its profit to owing roughly 3.5 years of it at close. The share half is printed: enough new stock that Element’s owners will hold roughly 44 percent of the combined company. If you owned a slice of Solstice in June, this deal hands nearly half of every future dollar of profit to somebody new, and it borrows three and a half years of earnings to do it. One more mechanical detail did quiet damage all week: the share ratio is fixed. Half a Solstice share is the payment whether the stock trades at $88 or $61, so every dollar the stock falls also shrinks what Element’s holders get paid, and hedge funds trade that shrinking gap mechanically, selling Solstice as they go. That is the arbitrage the CEO complained about on television. The skeptics also have real cracks to point at: the flagship segment’s profit margin fell 522 basis points (a basis point is a hundredth of a percent, so that is a drop of about 5 percentage points) in the first quarter, and 1,225 basis points, more than 12 full percentage points, in the quarter before that. Management’s explanation, in the filings, is product mix: the market is shifting from the old high-margin refrigerant gas to its replacement faster than the replacement’s economics have matured, and that shift more than offset favorable pricing. Put simply, a company eight months out of the wrapper asked the market for $14.5 billion of trust, paid with borrowed money and freshly printed shares. The market said no, erasing $4.1 billion of value in six sessions.

Why I think the selloff may be aiming at the wrong target. Everything the market punished last week sits on one side of the company: the deal, the debt, the dilution. The assets sit on the other side, and they did not change. The nuclear business grew 27 percent last quarter and already holds $2.2 billion of signed future orders, roughly five years of work at its recent pace of about $430 million a year in sales, at the only plant in the country that can do the job, while Washington orders more of exactly this capacity and shares the cost of building it. The Electronic Materials line grew 19 to 21 percent selling into semiconductor plants. These are the two fastest-growing things Solstice owns, and neither one appeared anywhere in July’s argument, which was entirely about the bill. So here is the hinge this edition turns on, in one question: can the boring cash engines, the refrigerants protected by a federal quota law and the uranium work already under contract, keep paying the new mortgage while the exciting semiconductor story proves itself out? If yes, the market just repriced the costume and forgot the cape. If no, then three and a half years of debt on a one-year-old company is exactly as dangerous as July says, and the sellers are right. There is also a wrinkle in how the nuclear profits get shared that I have not seen anyone mention, and it cuts against the easy bull case. Panel five has more than one way to end, and I will draw it honestly.

What could turn it. The deal proxy (the S-4, the filing that must precede the shareholder votes) will publish the boards’ projections and bankers’ math within months. Deleveraging is promised below 3 times within 18 months of closing, and the close itself is expected in the first half of 2027. Old nuclear contracts keep rolling onto today’s far higher conversion prices quarter by quarter. Management has committed $180 million a year of merger synergies by year three. And the company guided 2026 to $975 million to $1.025 billion of EBITDA before any of the deal lands, with the refrigerant quota law tightening the old gas’s supply on a schedule written into statute.

THE BUSINESS IN ONE LENS

Every screener on earth files Solstice under specialty chemicals. Fair enough. But read the annual report end to end and a stranger pattern starts repeating, and once you see it, you cannot stop seeing it.

Almost nothing this company sells is protected mainly by chemistry. It is protected by permission.

The refrigerants business, the largest by far, lives inside a regulatory transition: governments in the US and Europe are steadily retiring the old generation of gases and steering the market toward the low-warming replacements that Solstice builds and brands. The 10-K says the quiet part in plain text: these markets provide “regulatory protection for incumbent producers.” The nuclear business runs the one facility in the country with the license and the capability to do its job, and the same 10-K states that in a single sentence: “the sole domestic facility producing uranium hexafluoride gas.” Even the smaller lines live behind walls. The healthcare film is being joined by a medical-grade propellant certified for inhalers. The fiber sells into government armor programs. None of this is shampoo. You cannot decide to compete with most of this company by raising money and building a factory, because the factory is not the barrier. The paperwork is.

Honeywell did not spin off a chemicals company. It spun off a collection of permissions. And permissions are the hardest thing on earth to replicate, which is precisely why the market’s five-day verdict deserves a second look.

Now put numbers on the lens, because a lens without numbers is a slogan. Solstice reported $3,886 million of sales last year, split into two segments but really three engines.

Engine one, the transition machine. Refrigerants did $1,511 million of sales last year, up 16 percent, and grew 19 percent again in the first quarter. Here is the tension that will decide the hinge: the same regulation that guarantees demand for the new gas is currently squeezing profitability, because the market is shifting toward the replacement faster than its economics have matured. Management calls this product mix. The segment’s margin fell from 38.9 percent to 35.2 percent last year, and the market reads that slide as decay. Whether it is decay or a toll paid once is, I think, the single most important question in the stock, and it gets its full treatment when we reach the hinge.

Engine two, the tollbooth in Superman’s town. Nuclear did $356 million of sales last year, and here honesty requires an ugly detail the bulls skip: that was DOWN from $446 million the year before, because this is a business of lumpy multi-year contracts and delivery schedules, not a smooth subscription. The momentum now runs the other way, 27 percent growth last quarter, $2.2 billion of signed orders, and a Department of Energy agreement to share the cost of expanding the plant. One more detail, and hold it for the payoff later: Solstice does not pocket this tollbooth alone. The plant sells every kilogram through a joint venture, half-owned by a partner almost nobody discussing this stock has mentioned. What that does to the easy math is a payoff that comes after the paywall, and it is worth the wait.

Engine three, the quiet growth line. Electronic Materials did $409 million of sales last year and grew 19 to 21 percent in the recent quarters, selling ultra-pure materials into semiconductor plants, the end of the AI buildout where money is being spent fastest. (I rebuilt that spending from the hyperscalers’ own filings in is big tech’s AI capex mispriced; this is the shovel-seller’s corner of that story.) It sits inside the second segment with research chemicals and the armor fiber, which together add another $687 million.

Step back from the three engines and read the whole machine in one breath. The first segment earned $981 million of EBITDA at a 35 percent margin. The second earned $203 million at 18.5. Corporate costs ate $184 million, and the company poured $408 million into capital spending, up 38 percent, most of it aimed at the nuclear expansion and the growth lines. The balance sheet, before the deal: $2.0 billion of debt against $642 million of cash, about 1.4 times EBITDA. Modest. Boring, even.

Which brings us to the honest question this lens leaves hanging. If the permissions are so valuable, why did profit fall while revenue rose for a fourth straight year? The answer is not one thing, and the market’s five-day verdict treated it as one thing. Untangling it is the next section, because the bear case here is real, filed, and worth respecting before I argue with it.

WHY IT’S DOWN

The bears did not invent this selloff, and pretending otherwise would insult both them and you. Here is their case at full strength, every piece of it from the filings.

The earnings a screener sees are genuinely ugly

Net income fell 53 percent last year, from $605 million to $285 million, and the share attributable to Solstice’s own holders fell 60 percent. The filings show exactly where it went: transaction-related costs more than quadrupled to $117 million, the effective tax rate jumped from 24 percent to 56 percent in the spin year, and interest costs doubled. Some of that is the one-time price of being born as a public company. Not all of it. And there is a quieter line in the same table that should bother bulls more than the tax rate: last year’s 3 percent revenue growth came entirely from price and currency. Volume contributed zero.

The flagship’s margin is walking downhill in the middle of its transition

The refrigerants segment earned a 39 percent margin two years ago, 35 percent last year, 34.1 percent in the first quarter, and the fourth quarter of 2025 printed a 12-point year-over-year drop. Management’s filed explanation is product mix: customers are moving to the new low-warming gas faster than its economics have matured. The bear translation is blunter. The old gas was the profit engine, regulation is retiring it on a schedule, and the replacement has not yet proven it can earn what the original earned. A transition that guarantees demand does not guarantee margin.

The crown jewel just had a down year

The nuclear line everyone got excited about in June did $356 million of revenue in 2025, down from $446 million in 2024. Multi-year contracts deliver in lumps, and a down year proves nothing by itself, but it kills the tidy story that this asset only goes up. Whoever buys the tollbooth story has to hold it through delivery schedules that wobble.

The passengers are heavier than they look

The two growing engines carry the rest of the train. Healthcare Packaging fell 25 percent in the fourth quarter, which the filings attribute to customers destocking and the factory under-absorption that comes with it (a plant built for full production gets more expensive per unit when volumes drop). Building Solutions & Intermediates, the construction-linked line, fell 5 percent in the fourth quarter and 8 percent again in the first. Corporate costs rose to $184 million. None of these is a thesis-breaker alone. Together they mean the growth engines are pulling passengers, and the market noticed.

The screeners never stood a chance

Here is a mechanical reason the selling fed on itself. On last year’s reported earnings of $1.49 a share, June’s peak price was nearly 60 times earnings, so every quantitative screen on Wall Street filed Solstice under “wildly expensive chemicals company.” That earnings figure is polluted by the spin: the one-time transaction costs and the 56 percent tax year sit inside it. On management’s own adjusted guidance of $2.45 to $2.75 for this year, the price as I write is closer to 24 times. But a screen does not read footnotes, a one-year-old company has no long record to overrule the screen, and when the deal headline hit, the machines had no reason to hesitate.

A fish swallowing a fish its own size

Step back to the simplest number of all. On the day before the announcement, all of Solstice’s shares together were worth about $12.7 billion. The deal it signed is $14.5 billion including assumed debt. This company is buying something priced at more than the entirety of itself, and the promised reward, $180 million of annual synergies by year three, works out to about 1.2 percent of the purchase price. Deals this large relative to the buyer leave no room for integration error, and the buyer has been integrating itself for only eight months.

The seller prepared for this sale

Element Solutions is not a distressed seller. Its founder and chairman, Sir Martin Franklin, one of the most experienced dealmakers in the industry, signed a voting agreement supporting the deal the day it was announced. Five months earlier, Element had quietly upsized its own credit facilities, adding $450 million of term loans and a new $500 million revolver. None of this is improper, and none of it is comforting either: an eight-month-old management team is buying what a thirty-year dealmaker is selling, at a 15 percent premium.

And Solstice must close, whatever the weather

Read the merger 8-K closely and one absence matters: financing is not a condition of the deal. The $4.7 billion bridge is a 364-day loan, the permanent replacement will be a term loan and bonds at a rating the company itself describes as sub-investment grade, and if credit markets are ugly when the bill comes due, Solstice’s obligation to close does not care. The deal’s end date runs to July 2027, extendable into 2028, which means the arbitrage pressure on the stock has a very long runway. The proxy statement with the boards’ own projections has not even been filed yet. The market is pricing a deal whose interior nobody outside the boardrooms has seen.

Add it up and this is a coherent, filed, honest bear case: earnings down, margins compressing, the star asset lumpy, the smaller lines bruised, a deal bigger than the buyer, a shrewd seller, a must-close obligation, and a trust deficit that a company eight months old has not had time to earn its way out of. The market did not panic over nothing.

But notice what the market did NOT have during those five sessions: the paperwork underneath. The proxy with the boards’ projections does not exist yet. The merger agreement’s fine print, where that $513 million number from the top of this piece lives, went largely unread. And buried in the annual report’s investment note is something I have not seen in a single writeup of this company: a second quiet stakeholder, inside the nuclear business itself, collecting a share of the profits everyone spent June celebrating. The bears priced what they could see. The rest of this edition is about what they could not.

The paid half starts here. In it: the hinge, whether the quota-protected refrigerant engine and the uranium tollbooth can pay for this deal. The fine print, from the $513 million tripwire to the uncounted partner inside the nuclear profits. What I think the stock is worth, in three honest scenarios with the method shown. And where I land, as a real decision in the live model portfolio that comes with every dive for paid subscribers. Position one was Accenture. This edition decides position two.

Read more

In February I wrote that the traditional IT middleman was dead.

I meant it. That edition argued AI was collapsing the cost of building software to almost nothing, and that the people who bill by the hour to build it for you were standing exactly where the flood would hit first. Within a week the market agreed, loudly, and billable-hour businesses everywhere repriced. (If you missed it, the argument lives here: the traditional IT middleman is dead.)

Then, on the 22nd of June, I placed a buy order for the biggest billable-hours firm on the planet.

Accenture. Nearly eight hundred thousand people. The firm that boards call when they want change made without their own fingerprints on it. The stock had been cut in half in twelve months while everything else with AI in the story made new highs, and somewhere in that gap I stopped seeing an AI casualty and started seeing a mispricing.

Now, I usually never buy a position in one go. I accumulate. Small orders, spaced out, so the market has time to prove me wrong before I am fully in. What I placed that morning was the first of those small orders. A toe in the water, nothing more.

Here is the confession this edition is built on.

While preparing this deep dive, I opened my account to check my cost basis. There was no cost basis to check. The order never went through, and I had spent two weeks believing my accumulation had started when it had not. The stock is up around ten percent since. That stings more than a loss would. A loss at least means you acted.

So I am not writing this to defend a position. I do not have one. I am writing it to decide, with the filings open and you watching, whether that first order deserves to be placed again. And I do not get to shrug my way to an answer. Both of my bets are already on the table. I wrote the bear case in February. I placed the bull order in June. One of those two people was wrong, and this edition finds out which one.

Before we open a single filing, you should know something about this company. The market has sentenced Accenture to death twice before.

The first sentence came from its own family. Accenture began as the consulting arm of Arthur Andersen, the accounting giant, and by the late nineties the consultants were the ones making the money, handing up to fifteen percent of their profits to the parent nearly every year. They wanted out. The divorce took years of arbitration, and in August 2000 the ruling came down. The parent had demanded roughly fourteen and a half billion dollars. The arbitrator awarded it no damages at all. Just the roughly one billion dollars of profit-sharing money it was owed anyway, half of it already sitting in escrow, and one condition: surrender the Andersen name entirely, and go. An employee in the firm’s Oslo office named Kim Petersen won the internal naming contest with a made-up word, accent on the future, and in January 2001 Andersen Consulting became Accenture.

Within a year, the name it had been forced to surrender was radioactive. Arthur Andersen was Enron’s auditor, and it did not survive the scandal. Almost a century of reputation, gone. The orphan that paid a billion dollars to lose its family name had bought, by pure accident, the cheapest insurance in corporate history. It listed at $14.50 a share in July 2001 and spent the next two decades compounding into the largest technology consultancy on earth.

The second death sentence came a decade ago, and it will sound familiar. Cloud software was going to make integrators obsolete. Why pay consultants by the hour when the solution arrives turn-key, out of the box? You know how that ended. In the decade that followed, revenue more than doubled.

Now the third sentence has been handed down, and this time the executioner is supposed to be AI. AI writes the code, drafts the deck, does the work. Who needs eight hundred thousand consultants? The market believes AI is the largest business transformation of our lifetime, and it also believes the firm hired to run that transformation is dying of it. Both of those cannot be fully true at the same time.

Maybe the third sentence is the one that sticks. Companies do not get infinite lives, and sometimes the scare everyone laughed off twice is real the third time. That is exactly what this deep dive exists to find out.

I pointed Claude Code at every annual report, quarterly filing and earnings release Accenture has put on EDGAR since 2023. Eleven documents, every number pulled straight from the source, and one standing instruction: build the case against me.

(The whole rig, Claude Code running inside Obsidian with every filing it pulls saved into the vault, is its own edition: how I set up Claude Code as my investment research analyst, and the 2.0 rebuild that moved everything inside Obsidian. Start there if you want to run this yourself.)

What came back is stranger than the bull story or the bear story. There is a number Accenture spent two fiscal years leading its earnings releases with, and this year it quietly stopped appearing.

Hold that. We will get to it.

THE 60-SECOND THESIS

What they do. The world’s largest technology consultancy. 799,000 people, split almost exactly in half between designing large corporate transformations (consulting) and running the systems those transformations produce (managed services), on $69.7 billion of FY2025 revenue. Almost none of it needs capital: $600 million of capex against $10.9 billion of free cash flow last year. The asset is people billed by the hour. Which is exactly what the market has turned against.

Why it is down. The market has repriced Accenture as an AI casualty, and the dents are real, not imagined. Local-currency growth slowed 5, then 4, then 3 percent through FY2026. The US federal business, 8 percent of revenue, is being cut under DOGE (the US government cost-cutting department) and costs about a point of growth. Managed services bookings (new contracts signed, the order book), the bigger half of the business, fell 16 percent last quarter. Even the filing language has cooled: pricing went from “improved in several areas” to “relatively stable”.

Why I think that is the mispricing. Cut in half in a year, Accenture now trades at a free cash flow yield near 13 percent on management’s own $10.8 to 11.5 billion guide, by my math. The company is priced for decay and guiding for growth. One of those two is wrong. My view, and the reason for that June order: in the near term, everyone wants AI and almost nobody can implement it. Accenture has logged 104 quarterly client bookings of $100 million or more so far this year, up 13 percent, and its filings say clients keep prioritizing large-scale transformations to become AI-ready. Structurally, I believe companies hire Accenture for reasons AI does not remove: cover for hard decisions, someone else to blame, and speed they do not have in-house. That is an opinion, not a fact. The rest of this piece exists to test it.

What could turn it. The federal drag stops being news, since the guide already implies 4 to 5 percent growth without it. Capital returns the price ignores: a buyback raised to $7.5 billion, a dividend up 10 percent, $5 billion of net cash still on the books. And AI bookings visibly converting into revenue. Whether AI replaces the billable hour faster than it deflates it is the question that decides everything, and there is one number, quietly missing from the last two earnings releases, that sits at the center of it. Both get their full sections below.

THE BUSINESS IN ONE LENS

Most coverage treats Accenture as one giant consulting firm. The filings describe something more interesting: two different businesses wearing one badge, and AI is hitting them in opposite ways.

The first business is consulting. Senior people who design large transformations, project by project. That was $35.1 billion of revenue last year, growing 5 percent in local currency. The second business is managed services: multi-year contracts to run the systems and operations those transformations leave behind. That was $34.6 billion, growing 9 percent. And in the first nine months of this fiscal year, the second business became the bigger one for the first time, $27.9 billion against $27.6 billion. I had Claude Code pull the revenue tables from the last three 10-Qs and reconcile the two halves quarter by quarter, and that crossover is the quietest important fact in the filings.

Why it matters: consulting is where an AI wave shows up first as demand. Someone has to design the transformation, pick the models, rewire the workflows. Managed services is where AI shows up as deflation. It is the running of systems, the processing of invoices and claims and tickets, the recurring human hours that agents are supposed to replace. Same company, opposite exposure. You are not buying one bet on AI. You are buying both sides of it, in one stock.

And in the latest quarter, the two halves diverged exactly along that line. Consulting bookings rose 11 percent in local currency. Managed services bookings fell 16. One honest caveat before anyone extrapolates a single quarter: bookings are lumpy, and a handful of mega-deals moving by a few weeks can swing that number. But the direction matches the exact fear the market is pricing, and it is the line I will be watching hardest from here.

Step back and the machine itself looks nothing like a company in decline.

Margins tell the honest version: operating margin was 14.7 percent last year on a GAAP basis, 15.6 adjusted, and gross margin actually slipped to 31.9 percent, which the 10-K attributes to higher payroll costs. This is not a business getting more profitable per hour. It is a business getting bigger while paying more for its people. Hold that tension; it comes back in the hinge.

Geographically, the Americas are half the company, EMEA about a third, and Asia Pacific, the smallest slice, is suddenly the fastest growing at 9 percent. The balance sheet fits in one line: $10.2 billion of cash against $5.1 billion of debt, the first term debt in the company’s history, which is its own story for later. (Reading what sits above the equity holder is a discipline with its own edition: the whole debt structure from one filing.)

This is the machine the Andersen orphan built: capital-light, cash-gushing, buying a handful of small firms every year as its R&D department, and handing everything left over back to shareholders. The question this piece exists to answer is not whether the machine works. It is whether AI changes what the machine gets paid for.

WHY IT’S DOWN

The company admits the bear case in writing

I did not find the strongest version of the bear case in a short report or on a podcast. I found it in Accenture’s own annual report. I had Claude Code read the FY2025 10-K’s risk factors and pull every sentence about AI, and this one deserves to be read slowly:

“Some services and tasks currently performed by our people have been and will continue to be replaced by automation, including AI-enabled solutions, which will lead to reduced demand for our services and/or adversely affect the utilisation rate of our professionals, if demand for those services is not replaced by demand for new solutions and services... If we are unable to introduce or if our clients do not accept new pricing or commercial models that reflect the value of these AI-enabled solutions, our results of operations may be adversely affected.”

Not “may be replaced.” Will, unless new demand replaces what automation takes. The company that sells the AI transition has told its own shareholders, in a filed document, that AI shrinks the thing it currently bills for, and that survival means selling something new in its place. The whole stock is a bet on that unless. Everything else in this section is that sentence showing up in the numbers.

The growth is slower than the headline

Reported revenue grew 6, 8, and 6 percent across the three quarters of FY2026. Strip out currency and it is 5, then 4, then 3. The dollar did the flattering: foreign exchange added as much as 4.4 points in a single quarter. Management’s own full-year guide has narrowed from 2 to 5 percent in local currency to 3 to 4, with the top end doing the falling.

Washington took a bite

Accenture Federal Services was 8 percent of revenue last year, and roughly a third of the Health and Public Service group. Under DOGE, the filings describe procurement delays, price reductions, contract scope cuts, and outright terminations, with the GSA instructing every federal agency to review its consulting contracts. The result: a segment that grew 6 percent last year is shrinking this year, and management pegs the drag at about one point of total company growth, roughly $700 million by my math. There is also an open Department of Justice matter at the federal unit; it belongs in the risk section, and I will come back to it there.

Share

The order book cooled exactly where it hurts

Full-year bookings actually fell 1 percent in FY2025. This year started strong, then Q3 brought the number that matters: total bookings down 3 percent in local currency, and managed services bookings, the bigger half of the business, down 16. And the filings’ language about pricing has been quietly walking backwards. I had Claude Code line up the same sentence in the MD&A (the management commentary section of a filing) across four consecutive filings: “improved in several areas” became “improved in some areas” became “relatively stable.” That is how pricing pressure sounds when a company is choosing its words carefully.

The margin is being defended, not expanding

Gross margin fell 70 basis points last year to 31.9 percent, the payroll squeeze from the last section. The operating margin held up, but look at how: a $923 million severance and “business optimization” program that cut roughly 22,000 people from the February 2025 peak, though hiring has since resumed, and utilization (the share of employee time billed to clients) pushed to 93 percent, the highest level anywhere in these filings. The 10-K itself warns that utilization run too high damages engagement and drives attrition. Squeezing harder is a strategy with a ceiling.

And the challengers are smaller, faster, and sometimes your own client

The 10-K names the competition plainly: clients and ecosystem partners building their own AI capability, and “new AI-native companies.” The structural fear is simple. If AI collapses the cost of delivering technology work, the advantage of an 800,000-person pyramid built on billable hours inverts. Every mid-size AI-native shop is pitching exactly that story to Accenture’s clients right now.

Add it up and this is not a stock that fell on a misunderstanding. The bear case is coherent, filed, and partly visible in the numbers already. The market looked at all of it and cut the price in half.

The only question that matters is whether it stopped at the right floor. That question has one hinge, and the company’s own disclosures, including one number that quietly disappeared, are where we go next.

This is where the paid section begins. The free half gave you the setup, the history, and the honest bear case. The paid half takes the hinge apart, shows the number that vanished, and lands on what I am doing with that June order.

Read more

A better model shipped this month. Another one ships next month. You know the ritual. The timeline lights up, the benchmark charts fly, and somewhere in the back of your head a quiet worry starts. Is my whole setup obsolete again?

I want to offer you a different way to think about that worry, and it is one you already own. You are an investor. So put your setup on a balance sheet.

Every asset you have ever analysed sits in one of two buckets. Some assets depreciate. The computer you are reading this on lost value the day you bought it, and it has lost a little more every day since. And some assets compound. They grow the longer you hold them, quietly, and the growth feeds on itself.

Now look at where your attention goes when that worry starts. It goes to the model. And the model sits squarely in the first bucket. Whatever frontier model you run today will be mid-tier in six months, not because it got worse, but because something better always ships. That is fine. That is the one part of your stack you should want to be replaceable.

Here is the uncomfortable part. The second bucket, the compounding one, is where your actual edge lives. And for most researchers, it sits empty. Think about everything you have read this year. Every filing you pulled apart, every sharp thread you saved, every note you wrote on a company and swore you would come back to. Where is all of it right now? If you are honest, it evaporated. Scattered across bookmarks, chat histories, and folders you will never open again.

That is the real bottleneck, and it has nothing to do with which model is winning this month. The one asset in your research life that could compound is the one leaking value every single day, while all your attention goes to the part that was always going to be replaced.

The model is replaceable. The memory is not.

Most people have this allocation exactly backwards. All their attention goes to the depreciating asset, and the compounding one sits unbuilt.

For the last 120 days I have been building the compounding side, and this edition is the complete setup, start to finish. If you are new here, you do not need to read anything else first. Everything you need is on this page.

Some of you built the first version of this system with me, from How I Set Up Claude Code as My Investment Research Analyst. That piece is free, it is one of the most read things I have published, and most of it still holds. The installation steps, the walkthrough video, the foundations. But the setup did not stand still. I ran it every single day for four months, and it evolved. One tool got fired from the stack entirely. And a plain folder of text files quietly became the most valuable asset I own as a research operation.

Two months in, I want to know something, because this edition reads differently depending on your answer.

Whichever you picked, you're covered. If you built it, this is what yours grows into. If you dropped off or never started, this edition is the simpler second chance, and you'll see why in two sections.

The Claude Code + Obsidian Setup for Investment Research, From Zero

Let me build the whole thing in front of you, assuming you are starting from nothing.

The system runs on two tools. Both are free to install.

Claude Code is the analyst. It lives in a terminal, which sounds intimidating and stops being intimidating by day three. You talk to it in plain English. Tell it to pull a filing, read a transcript, screen for something specific, write a memo, and it goes and does the work on your computer, in your files, and shows you everything it did. If a terminal is a dealbreaker for you, hold that thought, because this edition removes that problem in a way that was not possible when I started.

Obsidian is the memory. It is a free app that points at an ordinary folder on your computer, and every note you create inside it is saved as a markdown file. Markdown, the .md format, is just plain text with simple symbols for structure, a # for a heading, a dash for a list. Humans can read it raw. And it happens to be the format AI models read most easily and most reliably, which is why almost every AI tool speaks it natively. So a folder of markdown files is, by accident of design, the most AI-ready research library that exists. No database, no export, no API. Obsidian’s only job is to make that folder pleasant to live in. (Obsidian setup guide)

Installing both takes about ten minutes. Obsidian is a normal app. Download it from obsidian.md, open it, and create a vault, which is nothing more than choosing a folder on your computer and giving it a name.

Claude Code installs with one command, and you should take that command from the source, because it stays current there. Open the official Claude Code quickstart at code.claude.com/docs/en/quickstart. You will see two install commands. On a Mac, copy the Terminal one and paste it into Terminal. On Windows, copy the PowerShell one and paste it into PowerShell. Press enter, wait a minute. Then type claude, press enter, and sign in with your Claude subscription. The analyst is installed. (Claude code setup guide)

And here is the standing rule for this entire edition. If any step confuses you, at any point, copy the confusing part, paste it into Claude, and ask it to walk you through it. You are setting up an AI analyst. Let it do its own tech support. This one habit is the difference between the people who finish setups like this and the people who quit at the first error message.

Analyst plus memory. That is the entire idea. Everything you read gets captured into the folder as markdown. Everything the analyst produces gets filed into the folder as markdown. Ask a question next month, next quarter, next year, and the analyst answers with everything the folder has ever collected behind it.

One more thing about this architecture, because it is the reason this edition will still make sense long after the current model leaderboard is forgotten. Nothing in this setup is married to Claude. Claude Code is the analyst I hire, but the chair itself is open. OpenAI ships Codex. xAI just shipped Grok Build. Moonshot has Kimi Code, MiniMax slots in too, and every serious lab now has a terminal agent of its own. Any one of them can sit in this same chair, pointed at this same vault, reading the same memory. The setup does not care whose model is winning the month.

That is why I will make a claim I almost never make about anything in AI. This setup will still be relevant six months from now, a year from now, probably two. And in this industry a year is like five anywhere else. It survives for one reason, and it is the whole argument of this edition: everything durable here lives in the folder, not in the model.

If you built the original version with me, you know there used to be a third tool sitting between these two. Cursor, the code editor I used as the workspace where the analyst and the memory met.

But run a system every day for four months and it tells you what it actually needs. Mine told me something I did not expect.

It did not need the middleman.

I Removed Cursor: Now the Entire Research Stack Lives Inside Obsidian

The middleman was doing two jobs.

Cursor gave me a terminal to run Claude Code in, and an editor to open my vault files in. That was it. Two windows in one app. And for months I treated that arrangement as the cost of doing business, until I watched what my hands were actually doing all day. Write in Obsidian. Switch to Cursor to run the analyst. Switch back to Obsidian to read what it produced. Switch to the browser to read something new. Clip it. Switch back. My research lived in one place and my attention lived in four.

Then I found out Obsidian could do both of Cursor’s jobs itself.

The first job, the terminal, is one community plugin. It is called Terminal, install it and a full terminal opens inside Obsidian, right next to your notes. In Obsidian, community plugins are built by users and installed from inside the app: Settings, then Community plugins, then Browse, and search for it by name.

Core plugins, which you will meet in a minute, are different, those are Obsidian's own, already built in, you just switch them on from Settings, then Core plugins. Install Terminal, open it next to your notes, type claude, and the analyst is now sitting inside the memory. Not next to it. Inside it. And notice what that plugin actually gives you: a terminal, not a Claude terminal. The day you want to run a different model’s CLI in there, Codex, Kimi, whatever ships next, it is the same plugin. Hold that thought, because it is the whole final section of this edition.

The second job, editing and reading the files, Obsidian was always better at. That is literally what it is for.

So Cursor left the stack. Not because it is a bad tool. Because the system stopped needing it, and every tool the system stops needing is one less thing a model upgrade can break.

And if the terminal itself is the thing that has kept you away from this entire setup, this is where I get to keep a promise from earlier. There is a plugin called Claudian. It puts Claude inside Obsidian as a chat window. No terminal, no black screen, no blinking cursor. You talk to it the way you talk to any chat, except it is sitting inside your vault, reading and writing your actual research files. The people who quit this setup at the terminal stage now have a door that does not require one.

Before I close the last window, meet the intake of the entire system: Obsidian Web Clipper. It is a free browser extension built by the Obsidian team, and it does one thing. Any page you are reading, an article, a filing, a thread worth keeping, one click, and it lands in your vault as a markdown file. In its settings you choose where clips land. Mine all go to one folder called raw, the system’s inbox, and that folder is about to matter in the next section. This is the moment reading stops being something that evaporates and starts being something you keep.

One more plugin closed the last window I was switching to. Web Viewer, and this one is not even a community plugin, it ships inside Obsidian. Turn it on and external links open inside Obsidian instead of your browser. Read the article there. Clip it from there with Web Clipper, and with Local Images Plus installed, every chart and image in the article downloads into the vault alongside the text, so the clip stays whole even if the page dies next year. Reading, clipping, filing, analyzing, writing. One window.

Two small quality-of-life touches before we move on. A plugin called Iconize lets you put icons on your folders, which sounds cosmetic and is, but a vault you enjoy looking at is a vault you actually use.

The second one I have never shown in writing before. At the bottom of my terminal sits a small status line: the folder Claude Code is working in, the model it is running with its context size, and a little rectangular bar showing how much of that context is already used, shifting from green to orange to red as it fills. One glance tells me where I am, what is running, and how much room is left before the model starts forgetting.

You do not write any code to get this. Claude Code has a /statusline command, and you describe what you want in plain English, the way you would describe it to a colleague. Mine was essentially: show me the current folder, the model name with its context size, and a small percentage bar for context usage that changes color as it fills up. Claude Code builds its own status line. It is a thirty-second setup, and it is the single thing visitors ask about first when they see my screen.

One last setting, and the screenshots have already given it away. At the bottom of my terminal it says bypass permissions on.

By default, Claude Code works like an advisory account. Before every single action, reading a file, writing a note, running a search, it stops and asks you first. Safe, but for real research it breaks the flow completely. You give it one task and then spend ten minutes clicking yes, yes, yes.

Bypass permissions turns the relationship discretionary. You give the mandate once, and it executes the whole workflow without stopping to ask. The community calls this YOLO mode, but discretionary is the honest word for what it is. Press shift and tab inside Claude Code to cycle between the modes, and the status line shows you which one is active.

One rule comes with the mandate, the same one that comes with any discretionary account: be careful what you hand over. A model with full permissions can edit and delete files. Point it at your research vault, never at your entire computer. The vault is its sandbox. Everything it needs lives there, and everything it can touch ends there.

That is the whole migration. Cursor out, one terminal and one browser in, both living inside Obsidian, and the count of windows I touch in a research day went from four to one.

But plugins are furniture. An empty vault with good furniture is still empty. Before this system can research anything, it needs a structure for information to flow through, and a brain that tells the analyst how to behave inside it. On day one, that is three folders and three small files.

CLAUDE.md: The One File That Makes Any AI Model Research Like You

Start with the structure. Open the terminal inside Obsidian, start Claude Code, and paste this:

Create this folder structure for my knowledge base:

- raw/ — this is my inbox where I'll dump source material
- wiki/ — this is your domain, you'll write and maintain everything here
  - Create a _master-index.md inside wiki/ with the heading "Knowledge Base Index"
    and a note that says "Topics will be listed here as they are created."
- output/ — this is where query results and reports go

Just create the folders and that one file. Nothing else yet.

Ten seconds later, three folders sit in your sidebar. raw is the inbox, everything you want the system to learn from gets dumped there. wiki is the library the analyst builds and maintains out of it. output is where finished work lands. That is the entire flow of information: in through raw, organized into wiki, produced into output.

But folders are just rooms. What makes this a research operation is the employee, and every employee needs to know the rules of the house. That is one file, and it is the most important file in the whole system.

CLAUDE.md sits at the root of the vault, next to the folders, and Claude Code reads it automatically at the start of every session. You write the rules once. It follows them forever. This is the same rulebook I published in the original setup edition, unchanged, because it has not needed to change in four months of daily use.

Create a file called CLAUDE.md and paste this in:

This file tells you how to maintain this knowledge base across every session.

## Knowledge Base Rules

- This is an LLM-maintained knowledge base. You are the librarian.
- The wiki/ folder is YOUR domain — you write and maintain everything in it.
  I rarely edit wiki files directly.
- raw/ is the inbox. When I dump files here, you process them into the wiki
  during a "compile" step.
- wiki/_master-index.md is the entry point. It lists every topic folder with
  a one-line description. Always keep this up to date.
- Each topic gets its own subfolder in wiki/ (e.g., wiki/ai-agents/) with its
  own _index.md that lists all articles in that topic with brief descriptions.
- Always use [[wiki links]] to connect related concepts across topics.
- When compiling raw material:
  1. Read the raw file
  2. Decide which topic it belongs to (or create a new one)
  3. Write a wiki article with key takeaways and relevant links
  4. Update that topic's _index.md
  5. Update wiki/_master-index.md
  6. If a raw file spans multiple topics, create articles in both and cross-link
- Keep articles concise — bullet points over paragraphs.
- Include a ## Key Takeaways section in every wiki article.
- output/ is for query results and generated reports.
- When answering questions, read _master-index.md first to navigate, then
  drill into the relevant topic _index.md, then read specific articles.
- When I ask you to "compile", process everything in raw/ that hasn't been
  compiled yet into the wiki.
- When I ask you to "audit" or "lint", review the wiki for inconsistencies,
  broken links, gaps, and suggest improvements.

(First published in the rulebook section of the original setup)

Read it once and you will see what it actually is. It is a job description. Claude Code is the librarian. The wiki is its domain, not yours. When you dump reading into raw and type compile, it files everything into topics, writes the summaries, and keeps the indexes current. You never organize, never tag, never remember where you put something. The discipline that kills every note-taking system ever built is delegated to an employee who never gets bored.

And here is the part that matters beyond any single tool. Notice the filename is just a convention. Claude reads CLAUDE.md. The other agents you met earlier read the same rules from a file called agents.md. The rulebook is plain markdown, it belongs to the vault, not to the model. Hold that, it is the entire final section.

Share

Two more small files live at my vault root, and they are optional, so one honest line on each. about-me tells the model who I am, my background, what I publish, who reads it. voice is my writing DNA, sentence structure, word choices, the patterns that make a draft sound like me instead of like a model. Neither is needed for research. Both exist because I also write a social media posts out of this vault, and the moment you ask the system to draft anything public, files like these are the difference between output you publish and output you rewrite. Build them if and when you need them.

That is the entire day-one setup. Three folders, one rulebook, maybe twenty minutes. It will look almost embarrassingly small. Good. Every compounding story does.

Here is what mine turned into.

What 120 Days of Daily Use Does to an Obsidian Research Vault

Every investor knows what compounding looks like on a chart. Almost nobody thinks about what it looks like in a folder.

Four months later, the structure has grown, and nobody designed it. Daily use forced every piece of it into existence, each folder earning its place by solving a problem I actually hit. That distinction matters. A structure you design up front is a guess. A structure that grows out of real research is a record of how the work actually flows. Here is the anatomy, in the order information moves through it.

Clippings is where reading enters the system. Every article, filing, and thread I clip from the browser lands here as a markdown file. This is the folder the Web Clipper from earlier feeds. Four months of reading that would have evaporated, sitting in one place, permanently searchable.

assets exists because of one plugin doing its quiet job. Local Images Plus catches every chart and image inside anything I clip and files it here. I have never manually saved a single file into this folder, and it is now one of the heaviest in the vault. Every chart from every article I have ever clipped, owned locally, alive even after the source page dies.

raw is the inbox for everything that is not a web clip. PDFs someone sends me, exported transcripts, notes I dump in a hurry. Anything that enters the system without a home starts here.

wiki is the processed library, the clean notes the system builds out of everything above. It is the one folder I mostly stay out of, for a reason that gets its own section in a moment.

output is where the analyst’s work lands, and it changed the most. In the beginning it was a loose pile of files. Now every piece of research gets its own project folder inside output, the data pulled for it, the working files, the final memo, all living together. One project, one folder, self-contained. If you have ever browsed a well-kept GitHub repository, that is the shape, except nobody codes anything and it never leaves your computer.

That is the anatomy. But a list of folders cannot show you the part that matters, the connections. For that, Obsidian has a built-in Graph View. Every note becomes a dot, every link between notes becomes a thread, and you see your entire research life as one living network. On day one, mine was three lonely dots.

Here is what it looks like today.

In above video, you can see the same vault in motion, panning from one node out to hundreds. This is what compounding looks like when it is your own reading.

Look at what is wired together in there. Published editions connected to the filings they were built on. Company notes connected to industry notes, connected to threads clipped months apart. None of these connections were planned. They accumulated, one reading session at a time, and now every new note lands in a web that already knows its neighbors.

This is the part nobody can hand you. A vault like this cannot be downloaded, templated, or bought in a zip file. It can only be grown. And that is exactly why it is the asset that compounds while the models come and go.

Which leaves exactly one question, and it is the one this entire edition exists to answer. You now have a growing library and a rulebook that thinks like you. But the analyst reading them today is one company’s model. What happens to all of this the day a better model ships from somewhere else?

Below the line, I show you the whole answer: the ten-minute migration that moves your entire research memory, the library, the rulebook, every project folder, onto any AI model that exists or ever will, without losing a single file. It is three steps. It is the reason I stopped caring about the model leaderboard. And once you have it, you will never rebuild your setup again.

Read more

On a broking desk, insider activity arrives as a feed.

A name buys. A green row appears. A name sells. A red row appears. Somebody glances at it, says “insider buying in so-and-so,” and the row scrolls away. That was the whole ritual. The feed was the truth, and the feed only spoke in buys and sells.

I did this for months before it occurred to me to ask where the feed came from.

It comes from a filing. Every time an insider does something with their stock, they have to tell the SEC, and it’s public within two days. The feeds are just those filings, cleaned up into a table you can scan in seconds. Who, how many shares, what price. Fast and easy.

But a table can only show you what fits in a column. And there is one kind of insider decision that doesn’t fit, so the table flattens it into the most boring row on the screen.

Picture a director who is owed money. Cash. Their retainer, their fees, sometimes their actual salary. Money the company already promised them. And instead of taking it, they tell the company to pay them in stock.

Think about what that is. Not a bonus. Not shares the board handed them as a reward. A person choosing to turn guaranteed money they were owed into stock they now have to hold and hope goes up. They are giving up a sure thing for a bet. On their own company.

On the feed, that decision looks like nothing. It doesn’t show up as a buy, because technically they didn’t buy anything, they got paid. So it lands in the same dull pile as every routine “company gave an executive some shares” event, the kind of row everyone is trained to skip. The choice, the part that makes it interesting, the fact that they chose this, is written one line down in a footnote. And the footnote is the part the table never prints.

So the signal isn’t hidden. It’s worse. It’s sitting in plain sight, disguised as the most ignorable row on the screen, and the proof that it matters is a sentence nobody scrolls down to read.

I covered the loud version of this a while back. Open-market buys, the green rows everyone can see, the insider spending real cash. This is the quiet version. The buy that doesn’t look like a buy.

My last edition argued that the edge in this job was never the tool. It was the judgment you wrap around it. This one is me proving it. One loop, built in the open, every number traced back to the filing it came from, every wrong turn left in.

And the lesson isn’t how to find these. Finding them is easy, a minute of work, and I’ll show you. The lesson is everything that comes after, because the moment you actually read what surfaces, you learn that almost all of it is noise. And the one that looks like the cleanest signal of all might be the biggest trap in the pile.

In fact, I already have one in mind. Look at this.

Why the Easy Part Is Worthless

Let me show you how easy the finding is, because it matters that you see it.

There is a search box on the SEC’s own website. It reads every filing ever submitted, including the footnotes, the part the feeds throw away. It’s free. No login, no terminal, no subscription. Anyone reading this can open it right now.

So I told Claude Code, in plain English, to go use it. Not with code. With a sentence.

Find me every insider filing in the last thirty days where someone wrote that they’re taking stock instead of cash. The phrases people actually use in these footnotes are predictable: “in lieu of cash,” “in lieu of director fees,” “elected to receive.” Search for those.

It came back with a pile. Around a hundred and sixty filings once the duplicates were stripped out.

And here is the trap, the one I want you to feel before I show you the way out of it.

A hundred and sixty hits feels like a result. It feels like you found something. Your instinct, my instinct for years, is to look at that number and think the work is done. You searched, the search worked, here’s your list. Most people would stop here, post the list, and call it a signal.

That instinct is exactly wrong, and it’s the whole reason most people get nothing out of insider data even though it’s free and sitting in front of them.

Because almost none of those hundred and sixty are what you’re looking for.

The search did its job. It found every filing with the right words in it. But the right words are not the right event. A footnote can contain the phrase “in lieu of cash” and still be describing something routine, something automatic, something that isn’t a choice at all. The words match. The meaning doesn’t.

This is the part no screen and no feed will ever do for you. A search can find words. It cannot read intent. It hands you a hundred and sixty filings and it has no idea that maybe five of them mean anything. Telling those five apart from the other hundred and fifty-five is not a search problem. It’s a judgment problem. It’s the thing the machine can’t do, which means it’s the only thing you’re actually being paid for.

So the real work was never the finding. It starts now, with throwing things away.

The Three Fakes That Look Exactly Like the Real Thing

When you actually read the filings the search hands you, they sort into a few piles. Most of them are fakes. Not fakes in the sense that anyone lied. Fakes in the sense that they contain the right words while describing something that means nothing about conviction. Learning to spot them on sight is the whole skill, so let me walk you through the three that fooled me first.

The first fake is the routine grant. A company pays its directors partly in stock every single year. It’s written into the policy. Nobody chose anything. On the day the grant lands, a footnote dutifully says the shares were issued in lieu of cash, because that’s the technically correct language for how the pay is structured. The words are there. But there was no decision. The director didn’t weigh a sure thing against a bet and pick the bet. They just received their normal pay in its normal form, the same as last year and the year before. A search can’t tell the difference between a choice and a routine. You can, the moment you read it.

The second fake is the whole-board sweep. This one is sneaky because it looks like a crowd. You’ll see eight directors of the same company all “electing” stock on the exact same day. Your gut says cluster. Eight insiders, same week, same direction, that’s supposed to be the strongest signal there is. But look closer and it’s the opposite. Eight people doing the identical thing on the identical day under the identical company program is not eight opinions. It’s one corporate event wearing eight nametags. The annual meeting happened, the standing policy kicked in, everyone’s pay converted at once. There’s no individual conviction in there to find. One event, not eight. The crowd is an illusion.

The third fake is the one that nearly got me, and it’s sharp enough that it deserves its own walk-through, so I’m giving it the next section. I’ll just name it here. It’s the election that says equity and pays cash. A footnote that reads like ownership, that uses every word you’d want to see, and then, buried in the same paragraph, quietly admits the whole thing settles in money. It is a paycheck in a costume. And if you’re skimming, you will count it as a real one every time.

Here’s what survives once those three piles are swept away. A real election has a person in it. Someone who was owed cash, who could have taken the cash, and who chose stock instead, in language that shows it was their choice and not the company’s policy. Not the whole board. Not every year on schedule. A specific human being deciding to turn guaranteed money into a bet on the thing they run or oversee.

And there’s one more filter, the one that turns a real election into an interesting one. The stock has to be down.

Think about why. A director electing stock in a year the share price already doubled is electing into a party. Easy to feel good, easy to look smart, the trend is your friend. That tells you very little. But a director electing stock while the price is falling, while the chart looks ugly and the easy move would be to just take the cash and wait, that’s a different posture entirely. That’s someone reaching for more of a thing the market is actively running away from. It still isn’t proof of anything. But of all the soft signals in this pile, election into weakness is the one that’s closest to conviction, because it’s the one that costs something to make.

So that’s the target. Elective, not routine. Individual, not a board sweep. Real stock, not a cash payout in disguise. And ideally, made into a falling price.

Now let me show you that third fake up close, because it’s the one that teaches you to never trust a footnote you only skimmed.

The Filing That Says Stock and Means Cash

Here is the one that nearly slipped through, and it’s the whole reason this is a filter and not a search.

I’m not reading a hundred and sixty filings by hand. That’s the point of doing this in Claude Code. The loop pulls them, opens each raw filing, reads the footnote, and applies the test I gave it. I told it the rule up front: an election only counts if it settles in actual stock the person has to hold. If the footnote says the thing pays out in cash, throw it away, no matter how much it looks like an election.

That rule is the entire reason a company called Enviri didn’t end up on my keep list.

On the surface, Enviri was a perfect match. A director, John Quinn. An election he made as a director. A clean acquisition, 3,155 units, a little over sixty thousand dollars. Every word the search was looking for was in that footnote. If all you did was search for the words and trust the hits, Enviri goes straight onto your list and you tell your readers a director showed conviction. A lot of people would stop exactly there.

But the rule made the loop read to the end of the sentence, and the end of the sentence is where it falls apart.

The security wasn’t called common stock. The filing named it, in parentheses, right at the top: a “Deferred Stock Unit (Cash).” And the footnote said what that parenthesis meant. Each unit “represents the right to receive the value, in cash, of 1 share” of the company’s stock.

Read that twice, because the trap lives in one clause.

He doesn’t get the share. He gets the value of the share. In cash. Later.

So nothing here is what it looks like. It looks like a director turning his pay into stock he has to hold. What it actually is, is a director parking his pay in an account that tracks the share price and pays out in money down the road. He owns no shares. He holds a cash claim whose size happens to move with the stock. If the price doubles his eventual cash doubles, if it halves his cash halves, but he never owns a thing and he never put guaranteed money at the risk that real ownership means. He stayed in cash the whole time. The cash is just deferred and index-linked.

That is not conviction. It’s a salary deferral with a market-linked twist, closer to a bonus structure than a purchase. And it filed under the same words, matched the same search, and would have sat in the same pile as the real elections, if the only thing I’d done was search.

This is the trap, and it’s the single most common way the signal gets faked. It fools you precisely because nobody lied. Every word in the footnote is true. The director did elect. The units do track the stock. The acquisition is real. The one thing missing is ownership, and it’s missing inside a clause most eyes slide past: “in cash.”

There’s a softer version worth seeing too, also caught by the same rule. A different company, Mosaic, had directors who elected to take half their units as shares and half “in the form of cash.” Same trick, smaller dose. Half of what reads as a stock election is just cash in disguise. One director took thirty-five percent in cash instead. So if you’d seen his filing and thought “nice, a director backing the company,” you’d have been a third wrong, and nothing on the surface would have told you. Only the last line of the footnote does.

Now here’s the part I actually want you to take with you.

I didn’t catch Enviri by being a careful reader. I caught it because I knew this trap existed before I ran anything, and I told the loop to kill it. Read every footnote to the end. If it pays in cash, throw it out. The machine did that a hundred and sixty times without skimming, without getting bored, without missing the clause on filing number ninety-three at eleven at night.

That’s the whole shift, and it’s why I keep saying the finding was never the job.

The reading is done by the machine now. You don’t sit there with a hundred and sixty filings and a coffee. What the machine can’t do is know what to throw away. It doesn’t know that “in cash” quietly kills the whole signal. It doesn’t know a deferred unit is a paycheck wearing a costume. You know that. And the moment you know it, you can teach it once and let it run forever.

So the edge isn’t reading carefully. Everyone can read. The edge is knowing what you’re reading for, and that part still lives in your head, not in the tool.

Let me show you what that looks like on a name I’m keeping.

The One I Kept, and Why I Almost Believed It

After the fakes are gone, you’re left with a short list. Real elections, real people, real stock that actually gets held. And one name on my list looked better than all the rest.

Agenus. A small biotech. The insider was Garo Armen, and he isn’t a junior board member. He’s the Chairman and the CEO. The man who runs the company.

And the footnote was the cleanest I’d seen in the whole run. Not “granted in lieu of cash,” the passive, company-driven language that makes you suspicious. This one read: “At his request and with the approval of the Compensation Committee, Garo H. Armen’s salary is being paid in stock, in lieu of cash.”

At his request.

Not the board’s policy. Not an automatic plan. His request. The chief executive asking to take his own paycheck in shares instead of money. And not once. He did it for the pay period ending in late May, then again two weeks later in June. A standing posture, repeated. The shares are fully vested the day they’re issued, so this isn’t a deferred unit that pays out later. He owns them now. Real stock, in his name, today.

And the stock is weak. It’s the chart I showed you at the start, the one I told you to hold onto. The five-dollar stock that gave it all back and kept sliding. This is the name. He’s been electing his salary into those shares the whole way down.

The rest of this one is for paid subscribers. The turn, what survived, and the claude code repo to run it yourself.

Read more

I have never written an opinion piece here. Not once. Every edition is a workflow you can copy and a repo you can tear apart line by line.

So when I break that rule, it is worth asking why.

I broke it because of the last twelve days, and because of a word the whole timeline has suddenly fallen in love with. Loop engineering. Prompting is dead, they say, the real skill now is the loop. I will come back to that word, because the people repeating it loudest are missing what it is actually for. Hold it in your head for now.

I spent the last twelve days inside eight investment firms and family offices. They invited me in, one after another, to see what a finance person actually does with AI. Real desks. People moving real money. I went into each room expecting to walk out with something to learn.

In all eight, I saw the same thing instead.

And it had almost nothing to do with prompts, or loops, or whatever the timeline is selling this week. It had everything to do with why almost all of that noise, prompting is dead, the labs are coming for your tools, hurry up and clone the right framework, is pointing you at the wrong target.

Let me take you into the rooms.

The Prototypes Were Already Gathering Dust

The first thing I noticed was how similar the eight rooms were.

Not the offices. The offices were nothing alike. Some were glass towers, some were quiet family setups with a few people and a great deal of money sitting behind them. What was identical was the AI.

Everywhere, it was doing the same two jobs. Summaries and slides. Feed it a long report, get back a shorter one. Feed it last quarter’s numbers, get back a cleaner pitch deck. A few people had a handful of prompts they were proud of, pulled out like party tricks. That was the ceiling. Useful, real, and about two inches deep.

Then there were the prototypes.

Almost every firm had one. Not built by them. Nobody on these desks writes software. At some point someone had gotten excited, found the budget, and brought in an outside engineer or an AI person to build the thing. A screening tool. A research dashboard. An internal assistant that was going to change how the whole team worked.

And in every case the thing was sitting there, open in a tab nobody clicked, gathering dust. The software worked. It ran. It did exactly what it had been told to do. It just did not do what these people actually do all day, and so nobody touched it.

I would ask about them, and the answer kept arriving in the same shape. *It is close. It is just not how we actually do it.*

That was the tell. The prototype was not bad. In most cases it was genuinely good. They had tried to pour their own framework into it, the way they actually screen, weigh, and decide, and somewhere between what they said and what the builder understood, it got lost in translation. What came back worked. It ran. It impressed everyone for a week. But it never became the thing they reached for when a real decision was on the table, and a tool that does not touch the actual decision is a fancy monument. It sits on the shelf and gathers dust.

Underneath all of it was a posture I started to recognize halfway through the second visit. Everyone was watching everyone else. Nobody wanted to go first. They had all heard the tools change every three months, so they were waiting, scouting, holding position, hoping someone down the street would prove it worked before they risked their own desk on it. The enthusiasm was real. The movement was zero.

And the people who wanted it most, the ones who pulled me aside with their eyes lit up, asking how to actually do this, were somehow the most stuck of all.

That part bothered me more than the dust. Because the dust already had an explanation, and it was sitting in the gap between the people who knew the work and the person who built the tool.

It Was Never an Engineering Problem

Sit with that gap for a second, because it is the whole thing.

On one side stood someone who could build software but had never sat in the chair. They had never carried a position overnight, never had to defend a number to a person who could lose money on it, never felt the difference between a screen that looks right and a screen you would actually act on. On the other side stood the investor who knew all of that in their bones and could not get it out in a form that survived the handoff.

And here is the part almost nobody sees. The investor was not being lazy or vague. The knowledge that runs a real desk is the hardest kind to put into words. A good analyst does not fully know their own process. Half of it is judgment they have never once had to explain, because they have never needed to. They just do it. Ask them to write it all down for a stranger to rebuild, and most of it does not come out. So the builder built from the half that could be spoken, and the half that actually mattered stayed in the investor’s head. That is the translation gap. Not a bad engineer. Not a bad investor. A thing that does not move cleanly from one person to the other.

Now watch what every firm reaches for next. A better engineer. A more senior AI person. More budget on the technical side. It feels right and it never works, because you cannot out-engineer a translation problem. The missing piece was never engineering. Pour ten times the technical skill into the same gap and all you get is a more impressive monument gathering the same dust.

The only thing that closes it is both understandings living inside one head. Someone who knows the work and can build the loop themselves, so there is no handoff for the judgment to die in.

That person barely exists yet. Not the engineer who picked up a little finance along the way. The investor who learned to build. It is going to be one of the most valuable seats in this business over the next few years, and I will come back to why, and what it means for you, before this letter is done.

You Have Been Asking Me for Something You Already Own

Which brings me back to the word I told you to hold.

Loop engineering. Strip away the breathless part and here is all it means. The machine does a step. It checks its own work. It throws out what fails the check. It comes back to you only when something is actually worth your time. Discover, verify, filter, repeat, with a gate at every handoff and a point where it stops. That is the whole idea. It is old. The label is new. Treating it like it arrived last week is how you end up paying a premium for something you could have built quietly months ago.

Now go back to those eight rooms with this in your hand. The reason none of those prototypes were loops is the same reason they gathered dust. A tool just runs. A loop decides. And a loop can only decide well when the person who knows what counts as a real signal, and what is noise to be thrown away, is the one wiring the checks. An outside builder cannot wire judgment they do not carry. So the firms got tools that run. Not loops that decide. The deciding part never made it across the gap.

Here is the uncomfortable part, and this one is aimed at the people in my own inbox.

You have watched me work this way for months. Whatever tool happened to be in front of me, the shape was always the same. [The screen that reads the filings, throws out the noise, and hands me the names actually worth reading]. [The vault that holds everything I have ever read, so nothing leaks and the research compounds instead of scattering]. None of it was a one-off trick. It was the same loop, over and over, long before anyone bothered to give it a name. I never called it loop engineering because it did not need one.

So when readers message me asking me to build them a loop engineering research workflow, I am not quite sure how to answer. It is already in the archive. You are asking me for something you already own.

Which leaves the only question that actually matters. If the loop is not the hard part, and the tool is not the edge, and the prompt everyone is now declaring dead was never the thing carrying the weight, then what is. That is the rest of this letter.

A Perfect Buffett Prompt Still Fails in Your Hands

Let me make the other side’s case first, because it is stronger than most people admit.

AI is going to eat the commodity layer of this work, all of it. The summaries. The screens. The extraction. Reading a three-hundred-page filing in a minute. Pulling a debt schedule, computing a fifty-year ratio, lining up ten comparables before you have finished your coffee. That is real, it is already here, and it is most of what those eight firms were using it for. Anyone who tells you AI will not flatten that layer is selling you something.

And the labs are not stopping at the inputs. They are now shipping the judgment layer too, or trying to. Pre-built financial-services agent templates aimed at banks, insurers and asset managers. Managed deployments inside the big consulting firms. The pitch is that you will not even need to build the loop, because a polished one will arrive in a box with your logo on it. So the honest version of the bear case is not just that the easy part gets commoditized. It is that the hard part is next.

Here is why I still think it is wrong.

Mohnish Pabrai built his entire public reputation on cloning. Copy the best investors, feel no shame about it, the idea does not have to be yours. He says it openly. And then, in the same breath, he draws a line most people skip right past. Once you see the idea, he says, you do all your own work, your own research, you make sure it is something you understand, you make sure it sits inside your circle of competence. Read that again. The most quotable champion of copying on earth is telling you the idea clones, and the work does not. You can lift the what. You cannot lift the who.

That is the whole thing. A perfect Buffett prompt is a cloned idea. Hand it to me and it does not become Buffett’s judgment, it becomes a very confident set of instructions pointed at a circle of competence I do not have. I wrote this once and I still believe it. Borrowing someone else’s investment skill is like wearing someone else’s prescription glasses. For two minutes you feel sophisticated. Then you realise you cannot see properly.

I have said this another way before, and I will say it again here because it is the hinge of the whole argument. Two successful investors can agree on eighty percent of the principles and still work completely differently, because the way they extract information, the way they connect dots, the questions they keep asking and the things they choose to ignore are never the same. You cannot inherit that from a zip file.

So when the labs ship you a research agent in a box, you are buying a cloned idea with a logo on it. It will be good. It will be fast. It will do the commodity layer beautifully. And it will work for everyone who buys it identically, which is exactly why it cannot be anyone’s edge. The moment a framework is packaged and handed out, its edge expires.

A quant I read said it cleanly: borrowed approaches expire when everyone else adopts them. The generative thing, the part that keeps producing new angles, is the part that was never in the box.

There is a second way frameworks fail, and it is quieter, because it is not about people at all. It is about time.

You have heard someone say the gold-to-silver ratio is high, so silver is cheap, sell when it reverts. They will hand you fifty years of data to prove it. But fifty years ago silver was mostly a monetary metal. What industrial demand it had ran through photographic film, and then digital cameras quietly erased most of that overnight. Today the demand has come back in a completely different form, wired into solar panels and electronics and EVs at a scale that did not exist when the ratio was first worshipped. The number still computes. It is just quietly measuring a thing that no longer exists in the same form. And I want to be careful here, because the failure is not using history. History gives you base rates and cycles, and the best investors lean on them hard. I lean on them so hard that when I built [my framework for researching any industry from scratch], the very first pillar, before growth, before unit economics, before a single number, was simply this:

how did we get here. The origin. You cannot read an industry until you understand the forces that shaped it. So history is not the enemy.

The failure is making it the final step. A fifty-year backtest is a cloned past, and the world it was cloned from changed underneath it while no one was looking.

Notice that AI is brilliant at exactly this layer. It will pull the fifty years and compute the ratio in seconds. Which is the entire point. If the machine can clone the past instantly, then the past was never the edge either. The edge is knowing when the regime broke. That is not in the data, and it is not in the prompt. It is in you.

So both clones fail the same way. The Buffett prompt is a framework that does not transfer across people. The gold-silver ratio is a framework that does not transfer across time. AI makes both clones free, and free is the tell. The thing it cannot copy, the judgment about which idea is real, which number still means what it used to, what to throw away, is the only thing left that was ever worth anything.

Which raises the obvious question. If you cannot buy the edge, and you cannot download it, and you cannot inherit it from a zip file, then where does it actually come from, and what are you supposed to do about it.

That is the part I have been building toward this whole time.

Build the Loop Around Yourself

So where does it come from.

It comes from the one place nobody can ship you. The work you have already done, the cycles you have already sat through, the questions you cannot stop asking, the things you instinctively ignore. Everything that got engraved into you slowly, over years, while you were not trying to turn it into a product. That is the circle of competence Pabrai was pointing at. It is not downloadable because it was never written down. It only exists as you.

The mistake almost everyone makes from here is to go looking for a better framework to copy. A sharper prompt. A more complete skill. Someone else’s system that finally works. I have watched smart people lose a year to this, hunting for the perfect borrowed process, and it never arrives, because the thing they are missing was never going to come from outside.

I have said this before, and I will keep saying it. A skill will not make you a better investor. Building your own, badly, watching it fail on a real thesis, and fixing it run by run, just might. Not because the file is valuable. Because by the third version, the clearest thing in the room is not the skill. It is your own process, finally written down where you can see it.

That is the real reason to build your own loop. Not because the loop is magic. Because building it is the only thing that forces the engraved, unspoken half of your judgment out of your head and into something you can run. The loop is just the excuse. You are the output.

And this is where the eight rooms come all the way back. The reason those prototypes gathered dust was that someone tried to write down a process they did not own. You will never have that problem with your own. You are not translating across a gap. There is no gap. The person who knows the work and the person building the loop are the same person, sitting in the same chair. Nothing gets lost, because nothing has to move.

That is the seat I told you we would come back to. It barely exists today. The investor who learned to build, who can take the judgment that lives only in their head and wire it into a loop that runs while they sleep. Not the engineer who studied some finance. The other direction. And I think over the next few years it becomes one of the most valuable profiles in this entire business, because it is the one profile that cannot be handed to anyone else.

Let me kill the obvious objection, because someone at one of those eight firms will raise it. The labs are coming for this too, they will say. They are already shipping research agents in a box, managed deployments, templates with your logo on them. Will they not just commoditize the build-your-own act as well.

No. They will commoditize the loop. They cannot commoditize what you point it at. A boxed agent is the most cloned idea of all, the same one sold to every firm on the street, which is exactly why none of them can build an edge on it. The lab can hand you a better engine every quarter. It cannot hand you the circle of competence to steer it. That part still has to be yours, and the better their engines get, the more the only scarce thing left is the judgment driving them.

There is one more reason this is worth starting now, and it is the part most people get backwards. The tools are not going to settle. They will get better every month, for years. Most people hear that and freeze, waiting for the dust to settle before they commit. But if the edge is the judgment you wrap around the loop, then a moving toolset is not a threat. It is leverage that compounds. Every upgrade makes your engraved judgment worth more, not less, because there is more horsepower behind the same irreplaceable hand on the wheel. The people waiting for it to stabilize are waiting for the one thing that gives them an advantage to disappear.

So stop hunting for someone else’s framework. Build your own loop, badly, around the one thing you already have that nobody can clone. Then let the tools race. You are the part that does not expire.

Enough Opinion. Here Is the Proof.

I have argued this long enough.

And argument is not what you come here for. You come here to watch the thing get built, then tear it apart yourself. So I am going to stop talking about loops and judgment and go do the only thing that has ever made this case for me. Build them in front of you.

This is the start of a series. Not a content calendar, a proof. Each edition is me taking some piece of judgment that lives only in my head and wiring it into a loop you can watch run, end to end, mistakes left in. You do not have to believe a word of the argument above. Watch me build, and decide for yourself whether the engraved part actually survives the trip into the machine.

Three are coming.

First, the insider signal your data feed deletes before you ever see it. Not the open-market buys everyone tracks. The quiet election, buried in a single Form 4 footnote, where a director quietly swaps guaranteed cash for stock. No vendor surfaces it. I will show you the loop that finds it, throws out the fakes that fool a naive screen, and hands back only the few worth a second look. Then I will tell you what those names did.

Second, Obsidian as the brain that any model can plug into. Not Claude only. Whatever runs best next month. The library sits still while the model behind it changes, and I will show you why the brain, not the model, is the part that compounds.

Third, the full loop engineering setup for investment research, start to finish. The thing readers keep asking me to build for them. The whole loop, in one place, so you can stop asking and start building your own.

I cannot hand you my loop. You already know that is the whole point. But I can build mine in the open, one edition at a time, and show you exactly where the judgment goes in. The rest is yours to engrave.

Before I Close

A quick, honest word about where the ideas in this letter came from, because the whole argument is that nobody does this alone.

The view that your edge is the slowly engraved version of you, not the tool in your hand, I have been circling for months in this newsletter. But the sharpest writers I read keep saying it better than I do, and the quants thinking out loud about loops put words to a thing I was already doing without a name. I steal from people smarter than me. That is not a confession. It is the method. The only original part is what survives my own pass.

Which is exactly what is coming, beyond the three editions above.

I am also starting something I have wanted to do for a long time. Real deep dives, on real companies, mostly the small and micro caps nobody else is bothering to cover. The kind of approach you already know, because it is the one I keep teaching. I find an idea from an investor sharper than me, on X, in a newsletter, in the rooms where good investors talk. Then I do the part most people skip. I run it through my own forensic process, AI doing the heavy lifting in the middle, and I find out whether the idea actually survives.

Because here is the belief under all of it. AI is not going anywhere. The only question left is how you use it. Almost everyone is using it to replace their judgment, let the model pick the stock, let the model tell them what to think. That is the fast way to be wrong with confidence. I use it the other way. To amplify what I can already do, not to do it for me. The view at the end is mine. The machine just got me there faster.

Idea to verdict, the whole thing in the open, a couple of times a month. Including the ones that fail my check, because a killed idea is often the more useful read.

None of this replaces the workflows. The frameworks are not going anywhere. This is on top.

So that is the plan. Less of me arguing, more of me building, and now some of me actually putting it to work on live names. If you have been reading this far, you are exactly who I write for.

Would you read it? Tell me. And if a sharp investor or a desk you respect should see this, send it their way. That is the only way this circle grows.

Disclaimer: Alpha with AI covers US-listed and other global securities for educational and research purposes only. Nothing here is investment, legal, or tax advice. AI-assisted analysis can be wrong, and every figure worth acting on should be checked against the primary filing yourself. Always do your own work and consult a qualified professional before making any investment decision.

Written by Shubham Borkar | Research & Insights by Shikshan Nivesh AI Team

Financial Clarity. Insightful Ideas.

Get 25% off for 1 year

We went paid three weeks ago. The response has been more than I expected.

Since then I have received messages I did not expect, not about the content, but about access.

A few days ago, a reader in Ghana wrote to me. They had been reading the free editions, found them extremely useful, and wanted to go paid. But the pricing page meant something very different in their economy than it does in New York or London.

It was not a complaint. It was one of the most thoughtful messages I have received since starting this newsletter.

It made me ask a question I should have asked on day one.

A Question I Should Have Asked on Day One

The question was simple.

Should access to quality research depend on where you live?

I already knew the answer. I just had not built around it yet.

So here is what we did.

If you are permanently located in a developing economy, including but not limited to India, Ghana, South Africa, Nigeria, Kenya, Brazil, Indonesia, the Philippines, or similar regions, just reply to this email, or write to us at info@shikshannivesh.com. Share your country, and we will work with you on adjusted pricing that fits your local purchasing power.

We will verify this through your payment details to keep the system fair for everyone. That is all it takes.

This is not a discount. It is what fair access looks like when your readers are spread across 50 countries, and a dollar is not worth the same in all of them.

Another Way In, and It Is Free

There is one more way to get full access, and it does not cost anything.

We just launched a referral program.

• Refer 7 readers, and you get one month of paid access.

• Refer 30, and you get three months.

• Refer 50, and you get six months.

If you know other investors or analysts who would find these editions useful, share your referral link with them. When they subscribe, the months are yours.

Terms apply, mostly to keep out spam and keep the program fair. But if you genuinely share work you believe in, this is the simplest way to earn full access without paying for it.

One More Way, If You Want Everything

And if you want the whole system, the Alpha with AI club premium membership now includes a complimentary subscription to this newsletter.

So premium members get the weekly playbooks, the community of 100s of global analysts, and every edition here, together.

Club Link

Four Ways In

When I started this newsletter, I wanted to make institutional-grade research accessible to people who were never handed it.

The message from Ghana reminded me that accessible has to mean accessible everywhere.

So there are four ways in now.

Upgrade at the regular price, the same one on the subscription page.

If you are based in one of the developing economies mentioned above, just reply to this email, or write to us at info@shikshannivesh.com, and we will adjust it for where you live.

Earn months of access by referring readers who would value this work.

Or join the Alpha with AI club premium and get this newsletter included with everything else.

The work behind the paywall is the same either way. The only thing that changed is who can reach it.

Disclaimer: Alpha with AI covers US-listed and other global securities for educational and research purposes only. Nothing here is investment, legal, or tax advice. Always verify against primary sources and consult a qualified professional before making investment decisions.

Written by Shubham Borkar | Research & Insights by Shikshan Nivesh AI Team

Financial Clarity. Insightful Ideas.

For years I traded off one number. The estimate.

Consensus said a company would earn this much. Analysts said buy, hold, or sell. I positioned around it, and so did everyone else on the floor. The entire job was reading what a room full of analysts thought a company was about to do.

It took me an embarrassingly long time to see the obvious problem.

The analysts were guessing. They had the same filings I had, the same models, the same blind spots. When they were wrong, they were all wrong together, and I was wrong right alongside them. I was spending my days reading the forecasts of people standing outside the company, trying to predict the inside.

Nobody around me was watching the one group that already knew the answer.

The company itself.

I learned this years later, on a broking desk in Bombay, from an older trader who barely looked at estimates. He watched what companies did after they reported, not what analysts predicted before. And he showed me a tell I have never been able to unsee since.

A company that just had a great quarter wants the whole world to look. The moment the numbers are out, it fills its calendar. Investor conferences, fireside chats, webcasts, roadshows, one after another for the next month and a half. It is throwing a party and sending invitations.

A company that just had a bad quarter does the opposite. It goes quiet. The calendar empties out. Fewer events, less marketing, a schedule with suddenly nothing on it. It is not hiding the numbers, those are already public. It is hiding from the conversation about them.

The company is voting on its own quarter with its own calendar. And that calendar is public, sitting on the investor relations page, weeks before any analyst updates a model.

Once you see it, you cannot stop seeing it. A full post-earnings calendar is confidence. A bare one is fear.

The best part is what comes next. This is not a read you do once, on one stock. It is a read you can run on every company you will ever hold, on every quarter, for as long as you invest. The same tell, the same hour after earnings, forever.

So I built it into something I can point at any ticker on demand. Take Lululemon. It just cut its guidance, the whole street turned against it, and Michael Burry started buying it anyway. I am going to run my whole read on it in front of you, live. I used Claude Code for this, but if that word makes you nervous, you can do the same thing in Claude Cowork. It is plain English the whole way down.

Why a Beat Throws a Party and a Miss Goes Dark

Think about what a company actually controls after it reports.

It does not control the numbers. Those are set the moment the quarter closes. It does not control what analysts write, or where the stock opens the next morning. All of that is out of its hands by the time the press release goes out.

But it controls one thing completely. Its own calendar.

In the weeks after earnings, management decides how visible it wants to be. It chooses whether to show up at the big bank’s tech conference. Whether to host a fireside chat. Whether the CEO sits for a webcast, whether the CFO goes on a roadshow, whether the investor relations team books anything at all. Every one of those is a choice, made by people who saw the full quarter before anyone outside the building did.

And here is the thing about people who feel good about their work. They want to be seen doing it.

A company that just beat, raised guidance, and likes its own story cannot wait to get in front of investors. It books everything. The four to six weeks after the print fill up with appearances, because every appearance is a chance to repeat a good story to the exact people who move the stock. A strong quarter wants an audience.

A company that just missed, or guided down, or knows the next quarter looks shaky, has the opposite instinct. Every investor event is now a room full of people asking the one question management does not want to answer. So the events quietly do not get booked. The calendar after a weak quarter is thin, not because nothing is happening, but because management has nothing it wants to stand up and defend.

This is not a theory about what companies should do. It is a pattern in what they actually do, and you can watch it happen in public. The investor relations page lists every scheduled event. The conference organizers publish their speaker lineups. None of it is hidden. It is just that almost nobody reads a calendar as a signal, because we are all trained to read the forecast instead.

That older trader’s whole edge was reading the thing everyone else ignored.

One honest caution before we go further, because I will not sell you a crystal ball. This is a tell, not a guarantee. A light calendar can mean a company in a quiet period, or one that simply does not market much. A packed one can be a coincidence of conference season. The signal is real, but it is one input, the same way insider buying is one input, the same way price action is one input. You stack it with everything else you know. Used that way, it is one of the cleanest free reads available to a retail investor, and until recently it was buried in work nobody wanted to do.

That last part is what changed.

The Four Things You Actually Read

The tell is simple. Reading it well comes down to four questions, asked in order.

One thing before the questions, so you know what you are looking for. This will not predict the quarter for you. What it does is catch the moments when a company’s behavior contradicts its own story, and that gap is usually the thing worth knowing.

One. How full is the calendar right now?

Go to the company’s investor relations page and look at the four to six weeks right after the earnings date. Count the scheduled events. Conferences, fireside chats, webcasts, an investor day, the non-deal roadshows that brokers host where management spends a day meeting investors without raising a dollar. Any appearance where management stands in front of the people who move the stock. That number, on its own, tells you something. A handful of events is a company that wants to be out there. A blank stretch is a company that does not.

Two. Is that more or less than the company’s own normal?

This is the move most people miss, and it is the one that matters most. A calendar with three events means nothing in isolation. Some companies are just busier than others, and many have a rhythm to their year, a heavy season and a quiet one. So you do not compare this quarter to last quarter and stop. You compare it to the same post-earnings window across the last four or five quarters. That history is the company’s baseline. Now this quarter has something to be measured against. If a company that reliably booked four or five appearances after every recent print suddenly has one, that drop is the signal, and you can trust it is a real change and not just the slow season. You are not reading the calendar. You are reading this quarter against everything the company normally does.

Three. What did management actually guide?

The calendar tells you whether management wants to be seen. This tells you what management said when it had to speak. On every earnings call, management gives a forward number. Revenue for next quarter, a full-year range, a margin target. That number is a promise, and a promise is only worth the track record behind it. So you do two things. You read the latest guidance, up, down, or held. Then you check the last several quarters of guidance against what the company actually delivered. A management team that quietly walks its full-year number down, after a history of hitting its targets, has just told you something the headline number did not. A team that always sandbags and then beats is playing a different game, and you read a soft guide from them differently. Guidance without the track record is noise. Guidance against the track record is signal.

Four. What does the street believe, and is anyone betting the other way?

The first three moves tell you what the company thinks of itself. This one tells you what everyone else thinks, so you can find the gap. Pull the current analyst picture. Count the buys, holds, and sells. Look at the average price target against where the stock trades now. You are not looking for the rating. You are looking for disagreement. A company going quiet on its calendar, after cutting its own guidance, while analysts have already abandoned it, is one kind of setup. The same company with one famous investor loudly buying into all of that is a far more interesting one. When the behavior, the guidance, the consensus, and a contrarian all point in different directions, you are not looking at a stock. You are looking at an argument. And arguments are where the money is.

That is the entire read. Calendar now, calendar against its own history, the guidance and the track record behind it, the consensus and whoever is fighting it. Four questions, every one of them answerable from public pages.

Share

Lululemon is sitting in exactly this kind of argument right now, which is why I picked it. It cut its full-year guidance and warned on the next quarter. The street has all but given up, has pointed out that of the thirty-two analysts covering it, only two rate it a buy, the most negative consensus he says he has seen for a company in this shape. And Burry whose Substack I read, has been buying it on the way down, calling the gloom overdone. published a deep-dive on it too, worth your time if you want the long version. Four signals, four directions, one stock. The perfect thing to point a fresh read at.

Here is the catch, and it is the only honest objection to any of this.

Done by hand, this is an afternoon. You are clicking through an investor relations site, copying event dates into a notepad, going back five quarters and doing it again, then reading four earnings calls for the guidance, then opening three more sites to rebuild the analyst picture. By the time you finish for one stock, the moment has cooled, and you have one stock. Nobody runs this on their whole watchlist by hand. That is exactly why the signal stays underused, sitting in plain sight, ignored because the reading of it is tedious.

So I stopped reading it by hand.

What follows is the same four questions, handed to an AI that does all of it in one pass, on a stock with the whole street on one side and Burry on the other, every number traced back to a public source you can check yourself.

The Whole Thing Is One Paragraph You Type in Plain English

This is the part where most newsletters hand you a six-hundred-word prompt with roles and rules and capital letters. I am not going to do that, because the entire reason to run this inside Claude Code is that you can just talk to it.

So here is the actual thing I typed. One paragraph, plain English, the way I would explain the job to a junior analyst sitting next to me. Read it once and you will see all four moves living inside it, in order, without a single bullet point.

Look at Lululemon, ticker LULU. It just reported and cut its full-year guidance. I want to know how the company is behaving after this quarter, because companies that feel good about a quarter fill their investor calendar and companies that don't go quiet. So first, go to its investor relations page and any conference schedules you can find, and list every investor event on the calendar in the four to six weeks after the earnings date. Then do the same thing for the last four or five quarters, the same window after each earnings date, so I can see whether this quarter is busier or quieter than the company's own normal. Next, read the latest earnings call and tell me exactly what management guided for next quarter and the full year, and compare that guidance to what they guided and then actually delivered over the last few quarters, so I know whether this is a team that keeps its word. Finally, pull the current analyst picture, how many buys, holds and sells, and the average price target versus where the stock trades now. Cite a public source for every number, and if anything isn't disclosed, say so instead of guessing. At the end, give me a plain-English read: is this company acting confident
or acting scared, and where does its behavior disagree with what the street believes?

That is it. No setup, no persona, no formatting instructions. You could read that paragraph out loud to a person and they would know exactly what to do. That is the whole point. The model does too.

Notice what is hiding inside it. The first sentence is move one, the calendar now. The second is move two, the calendar against its own history. The third is move three, the guidance and the track record behind it. The fourth is move four, the street and the gap. The four moves are the framework. The paragraph is just the framework spoken like a human.

And here is the part I want you to take even if you stop reading here and never open Claude Code. You do not need the tool to use the idea. You can paste that paragraph into plain Claude or any CLI you already use, change the ticker, and get a usable read on any company you hold. The framework travels. The tool only makes it faster and lets you point it at ten names instead of one.

When I ran this on Lululemon, here is what came back.

I Pointed It at Lululemon. It Told Me My Own Rule Was Too Simple.

Here is the part I did not plan for.

I expected a clean win. Lululemon had just cut guidance, the street had turned, the story looked broken. The tell predicts a company in that spot goes quiet, empties its calendar, hides from the conversation. I pointed the read at it expecting to watch the calendar thin out and prove the whole thing in one screenshot.

That is not what happened. And what happened instead is worth more than the clean win would have been.

First, the run. Three ways to do this yourself, pick the one that scares you least. Open the Claude desktop app and use Claude Code or Cowork, no terminal, paste and talk. Or run it the way I do, inside Obsidian, with the terminal plugin installed from community plugins so Claude Code works right next to your notes. I walked through that whole setup in how I set up Claude Code as my investment research analyst and the Obsidian layer in the part four edition. Or if you live in a terminal, use Codex or any CLI. The paragraph is identical everywhere.

I pasted the paragraph from the last section, the plain-English one, swapped in nothing, and let it work.

The entire run, start to finish, nothing cut. This is the whole process behind the read below. Watch it pull five quarters of investor events on its own.

It did not answer from memory. It fanned the job out, hit the investor relations page, the press releases, the conference schedules, the transcripts, and pulled five quarters of history before it said a word. The whole session is in the repo, every source link intact, so you can check any number I am about to show you.

Then the answer came back, and the first thing it told me was that my hypothesis ran straight into a structural fact. Lululemon barely uses its investor calendar in any quarter at all.

Move one and two, the calendar. Here is what it actually counted, the events in the four-to-six-week window after each of the last five earnings dates:

Read that and the naive version of my tell falls apart. The most recent quarter has two events, tied for the busiest of the five. If I had just counted events, I would have concluded Lululemon was feeling confident. The opposite of the truth.

The reason is the thing the model caught and I had not thought about. Some companies barely market to investors at all. Lululemon does not press-release its conference appearances. Its calendar is almost always near-empty, so there is nothing left to thin out. Run a raw event count on a company like this and the signal is just noise.

This is the moment the run earned its place in this edition. Because then it found the actual signal, the one hiding underneath the count.

The real tell was not how short the calendar got. It was what the company refused to add to it. A business down around 60 percent over the year, run by an interim team, fighting a proxy battle, did not schedule a single voluntary thing. No analyst day. No special investor update. No bonus conference circuit to stand up and defend the story. It did the legal minimum and nothing more. One of its two events this quarter was a mandatory annual meeting, and this year that meeting was a contested proxy fight tied to a Chip Wilson cooperation agreement. The company was forced into that room. It did not choose to enter it.

So the tell was right. My way of reading it was too crude. The company was hiding. It just did not need to empty a calendar to do it, because the calendar was already empty. You read the silence against the situation, not the event count against last quarter. That correction is the whole reason I am glad I ran this live instead of writing it from theory.

Move three, the guidance, and this is where the run got sharp. This is the move you would never get from counting anything. The model read the last five quarters of guidance against what the company actually delivered, and found a split that tells you exactly how much to trust management.

Every single quarter, for five quarters running, Lululemon beat the top of its own EPS guide. Five for five. This is a team that always clears the bar it sets for the next ninety days.

And across that same stretch, it cut its full-year guidance three separate times. Beat the quarter, lower the year. Beat the quarter, lower the year. The pattern in one line: they keep their word on the short bar they set low on purpose, while the real trajectory of the business grinds down underneath them.

That is the difference between a management read and a headline. The quarterly guide is a sandbag they will probably clear again next print. The full-year guide is the one quietly telling you the franchise is deteriorating. Trust the short word. Distrust the long one.

Below the line is where this resolves. The full analyst picture and the one tension buried inside it that I think matters more than the guidance cut itself. The exact point where management’s story and the market’s story split clean in two. What Michael Burry is seeing here that the entire sell-side is not. The plain-English verdict on whether this company is acting confident or acting scared. And the finished read you can point at any stock you hold, with the whole conversation and every sourced link in the repo.

This is where the paid section begins. The free half gave you the method and the full live run. The verdict, the analyst picture, and the saved skill are below for paid subscribers. If this is the kind of work you want every edition, consider subscribing.

Read more

Nobody teaches you what to do on the day your stock breaks.

There are thousands of newsletters about investing. I write one of them. We teach frameworks, checklists, workflows. All of it built on problems someone already solved.

Then a week like this one arrives.

Broadcom, one of the ten largest companies in America, is down 8.5 percent over the past month. That number sounds manageable until you see how it happened. Almost all of it came in the last five trading days, a drop of roughly 20 percent.

The stock spent most of the month going up and then lost a fifth of its value in under a week. The whole chip space went down with it. If you were holding it, you were not thinking about frameworks. You were staring at a red screen after lunch, asking the three questions every holder asks and no one answers in advance:

Is it my stock? Is it my sector? Or is the whole market breaking?

Here is the uncomfortable truth about investing content, including mine. We teach solved problems. Your portfolio produces unsolved ones.

That gap is not your fault. No writer can cover your problem before it happens. The market invents new ones faster than anyone can publish.

So this edition is different. I am not handing you a prompt for this situation. I am going to build the answer in front of you, the same evening the problem showed up, using a stock I had never researched in my life until this week. And then I am going to show you the part that actually matters: the method underneath, the one you will reuse next month on a problem nobody has written about yet.

Solved problems make you informed. Unsolved ones make you money. This is how you work on the unsolved kind.

I Already Wrote This Edition Once. It Could Not Help Me This Week.

In November I published an edition about exactly this situation. An Indian stock called Power Finance Corporation had dropped 8 percent while its sector barely moved, and I showed you how to diagnose whether the problem was the company, the sector, or the market. You can read the whole edition here.

That edition had a 600-word prompt. It asked you to download two CSV files from the Indian stock exchange first. Then you pasted everything into an AI and waited.

It worked. Readers ran it. I was proud of it.

This week I went back to it, and I saw it for what it was. A solved problem, frozen in place. The prompt knew the Indian market. It knew NSE data files. It knew nothing about Broadcom, US sector funds, or where American capital hides when it gets scared. My own prompt could not help me with my own framework.

That is the problem with prompts. A prompt is an answer. It answers the exact problem it was written for, and nothing else.

And investing problems do not repeat politely. They come back wearing a different ticker, in a different market, with the data living somewhere new. Whatever hits your portfolio next month will not match anyone’s template. Including mine.

So this time I am not handing you the answer. I am going to show you how the answer gets built. The whole thing took one evening.

I used Claude’s new model, Fable 5. You may have seen it called Mythos 5, it is the same model. And I ran it inside Claude Code, directly in the Claude desktop app. If you have the Claude app on your computer, you already have everything I used. No terminal. No installation. Open the app, and Claude Code sits right there next to Cowork. (If you want the heavier setup I personally run, terminal, Obsidian vault and all, that is its own edition here.)

What follows are the five instructions I typed, in order. Each one is plain English. Each one carries a principle you will reuse on problems I will never write about.

But first, the three questions everything hangs on:

1. Is my stock weak on its own?

2. Is my sector shifting?

3. Or is the whole market repricing?

Before I show you a single number, commit to an answer.

Share

The Build, In Five Moves

What follows is the entire session, in order. Five instructions, typed in plain English, inside Claude Code in the Claude desktop app. Each move carries a principle. The principles are the edition. Broadcom is just the example.

The whole session is also here as an 8-minute recording, uncut, nothing edited out. My suggestion: read the five moves first, then come back and watch it once you know what you are looking at. The moment the data source dies and the model reroutes itself is worth seeing with your own eyes.

Watch for: the model designing the data plan from three questions, the data source dying and the self-reroute, the patch on 48 files, the verdict, and the skill being written at the end.

Move One. State the problem, not a prompt.

In November I would have started by writing a prompt. Role, inputs, computation steps, output format, NA rules. Six hundred words before the AI did anything.

This time I typed this:

Tech is selling off this week and Broadcom has dropped around 20 percent in five days. I want to answer three questions: Is Broadcom weak on its own? Is the whole sector breaking down? And where is the money actually going? Before you do anything, tell me exactly what data you would need to answer these three questions properly, and where you would get it from.

Notice what that last sentence does. It does not ask for analysis. It asks for the data plan. Most people skip this and jump straight to the answer. The data plan is where the quality gets decided.

Full output:

data-plan.md

What came back was a better research spec than my November prompt. For question one, it wanted Broadcom’s prices and volumes, the same series for six direct peers, and the news flow around the drop. Then it added something I had not asked for: beta context. Broadcom moves more than its sector on a normal day, so a sector drop already explains part of any Broadcom drop. The real question is the residual, the part the sector cannot explain.

For question two, it refused to judge a sector by one stock. It wanted breadth, what fraction of the chip names are actually down, and it wanted equal-weight versus cap-weight tech, because a megacap problem and a sector problem look identical in the headline index.

For question three, it opened with a caveat I want to frame: every share sold is a share bought. “Where is the money going” is really two measurable things, what is catching a bid, and where fund flows land. It flagged in advance that hard flow data lags by days and everything before that is inference from prices.

My November prompt had none of this reasoning. It had my instructions, frozen. This had thinking, because I handed it questions instead of commands.

Move Two. Files, not memory.

The single biggest source of AI errors in finance is letting the model carry numbers in its head. Numbers read off webpages, half-remembered from training, blended across dates. So the second instruction set one rule for the rest of the session:

Good. Pull all of it. Two rules. First, everything lands as files in a dated data folder: daily closing prices and volumes as CSVs going back about three months for every ticker and ETF you listed, and the news timeline and any ETF flow findings as a markdown file with the source of every item. Second, from here on, every number we use must come from these files, not from memory or a web page.

Then I watched it work, and this is the part no demo video shows you, so I am leaving it in.

Its first-choice data source turned out to be blocked behind a JavaScript wall. It fell back to Yahoo Finance’s API, which rate-limited it. It fell back again, to the yfinance library, which worked. Then its integrity scan caught that Yahoo had not yet backfilled the previous day’s close, so the latest row of every file was empty. It patched all 48 US-listed series from a different endpoint, verified the dates, and documented the patch in a README. Then my Firecrawl ran out of credits mid news-pull and it switched to built-in web search without being told.

None of this was my plan. All of it is why the workflow survives contact with reality. Twenty minutes later the folder held 50 price files covering 64 trading days, a news timeline where every single item carries its source link, and a README documenting where every number came from, including the patch.

That folder is the product. Everything after this is just reading it. The full folder is in the repo:

repo link

Move Three. Models interpret. Code computes.

I had the third instruction ready in my notes. Here it is, exactly as I planned to type it:

Now compute from the files: returns over 1 day, 5 days, 1 month and 3 months for Broadcom, its sector benchmarks, all the major sector ETFs and the safe-haven assets. Also compute Broadcom’s return minus its sector’s return, 20-day volume z-scores for the sector’s top holdings, and sector breadth, meaning what percent of holdings are positive. Do all of it in code and save the result tables as files.

I never got to type it.

The moment the data folder was full, the model went ahead and computed all of it on its own, in code, and saved the tables next to the prices. Not because it read my mind. Because the rule I set in Move Two, every number must come from these files, left it no other way to proceed. Set the architecture right and the discipline follows without being asked.

In November, the AI did this arithmetic in its head, reading numbers off pages and calculating returns as a language model, which is exactly where language models are weakest. The wall of warnings in my old prompt existed to compensate for that. This time the arithmetic never touched the model. A script ran on the folder. Same input, same output, every time.

Here is what the files say, and I am giving you the exact figures because every one of them traces to a CSV in the repo:

Broadcom fell 18.6 percent over the five days. Its sector fund, SMH, fell 6.5 percent. Broadcom’s beta to its sector, measured on the three months before the event, is 0.76, which means the sector’s week only explains a 4.9 percent drop. The residual, the Broadcom-specific part, is roughly 13.6 percentage points. And it happened on volume running at 2.2 times the three-month average. That is not drift. That is conviction selling.

The sector table is just as blunt. 22 of 25 chip names down on the week, median minus 6.3 percent. Equal-weight tech fell as much as cap-weight tech, 8.9 versus 8.8 percent, so this was not just a megacap story. And yet: every benchmark, SMH, SOXX, XLK, QQQ, SPY, still sits above its 50-day average. Every one except Broadcom itself.

Move Three output key-stats.md

Below the line: the verdict on all three questions, the poll resolved, the moment the model audits my own rules and finds where they are blind, the fourth classification that did not exist this morning, where the money actually went, and the finished skill you can install and run on any stock you own.

Read more

A reader messaged me last week and asked if I’d break down the SpaceX IPO.

I almost said no. Not because it isn’t interesting. Because I don’t write about whatever is trending that week. Trends rot. In AI especially, the hot thing today is stale in three days, and I’m not interested in teaching you something with a shelf life. I write about workflows that still matter a year from now, and I usually find them buried in messages like this one, from readers telling me what they’re actually stuck on.

This one was worth it. Because SpaceX isn’t really the lesson. The lesson is the document.

I’ve admired this company for years. Not Musk the personality. The work ethic, the engineering DNA. There’s a story about how that DNA works that has stuck with me for a long time.

An engineer comes to Musk and says he needs to buy a part. It costs $1,200. Musk doesn’t ask why they need it. He asks what it’s made of. Some kind of metal, they say. He asks what that metal costs. They check the London Metal Exchange. Fifty dollars a kilo. He just looks at them. And eventually the room works out what he already knows, that the $1,200 part is maybe $300 of raw material and they should build it themselves.

That’s the whole company in one scene. He refuses to accept the price he’s handed. He breaks it down to the raw inputs and rebuilds the real number from scratch.

Every competitor knows this story. None of them have the DNA to actually do it. That refusal is why these companies trade at a premium nobody else gets, and why Wall Street bends the knee.

So here is what I find funny.

The man who has never accepted a sticker price in his life just handed the public a 277-page sticker price. The SpaceX prospectus. Plus a hundred-page addendum. Dozens of photos of rockets going into orbit. A soaring story. And somewhere inside all of it, the five or six numbers that actually decide whether this is the business of a generation or the most expensive story ever sold to retail.

Twenty years ago a prospectus was short. You could read one in an evening and know what you owned. Today they run two to three times longer, and not because there is two to three times more worth knowing. The length is the point. Noise is a great place to hide signal.

I’m not going to read 277 pages to find five. Not in 2026. Time is the only thing I can’t get back, and there is no reason left to spend it this way.

So I did to the prospectus what Musk would do to that $1,200 part. I refused the price. I broke the document down to its raw inputs and rebuilt the real number myself.

This is how. And it works on any prospectus, any annual report, any 300-page document built to wear you down, whether you’ve been investing for years or you’re about to analyse your first IPO.

One thing before we start. The title says Claude Code, and if that sounds like a developer tool you’re about to be locked out of, relax. There’s a no-terminal version of this exact workflow, same power, no commands, and I’ll show you both. Nobody gets left at the door.

DATA IS NOT INFORMATION

has a framing in his recent SpaceX note I keep coming back to. He calls it data versus information. The way I've come to interpret it for myself is simpler. Data is not information.

A prospectus is full of data. Page after page of it. Numbers, disclosures, risk factors, diagrams. What it is short on is information, the small number of things that actually change what you believe about the business. Damodaran has even written a paper on this. Prospectuses today run two to three times longer than they did three or four decades ago, and not because companies have three times more worth telling you. The extra length is mostly defensive. Risk factors written to be unfalsifiable. Language vetted so nothing in it can be used against the company later.

So before I open any filing, I do something that sounds backwards. I read what other people wrote about it first.

Here is why. I love reading. I will happily spend an evening with a great piece of writing about a business. But there is a difference between reading an author and reading a document. An author I respect is accountable to their own reputation. They have a name on the page and a track record you can check. A prospectus is written by lawyers who are accountable to one thing, protecting the company that pays them. Given the choice, I will read the author first, every time. Not because the author is always right. Because the author is trying to inform me, and the document is trying to sell me.

So that is where I started with SpaceX.

READ THE AUTHORS BEFORE THE DOCUMENT

I started with Aswath Damodaran.

He had just published a post-prospectus update on SpaceX, revisiting a valuation he’d first attempted before the filing came out. This is exactly the kind of reading I trust. A man who has valued more companies than almost anyone alive, showing his work, telling you what he got wrong the first time and what the prospectus changed. I read it once for the story and once for the numbers. Then I clipped the whole thing into my Obsidian vault so it sits in my own research library, searchable, permanent, mine. If you want the setup that does this, I’ve written about how my research library works before which you can read here.

Then I went to X.

Not for hot takes. X has the largest live investing community on earth, and buried in the noise are people who have actually done the work and published their reasoning. I asked Grok to pull me the most substantive bull cases and, just as deliberately, the most substantive bear cases, with names and links so I could read the originals myself. I did not want a summary that hid the disagreement. I wanted the strongest version of both sides. I clipped the results into Obsidian alongside the Damodaran piece.

A note on this, because it matters. Everything I gathered at this stage is opinion. Smart opinion, sometimes. But opinion. None of it is a fact until I have checked it against the filing. A bull who tells you Starlink subscribers doubled is giving you a lead, not a number. The number lives in the prospectus, and that is the only place it counts.

Now I had a folder full of intelligent people arguing about SpaceX. And only then did I open the prospectus itself.

That is the order that matters. Authors first, to build the story and the questions. The document last, to settle them.

THEN I OPENED THE PROSPECTUS

So I opened it. Two hundred and seventy-seven pages, plus a hundred-page addendum. Dozens of photographs of rockets leaving the atmosphere. A story about humanity becoming multiplanetary, written to make you feel something before you’ve read a single financial statement.

I read enough to confirm what Damodaran already warned me about. The signal is in there. It is just buried under a great deal of material that exists to fill pages, not to inform a decision. Risk factors that could apply to any company on earth. Sentences engineered by lawyers to mean as little as possible while technically disclosing something.

And here is where I made my decision. I was not going to read all of it.

Not because I couldn’t. I’ve read plenty about this company over the years and enjoyed every word of it. But there is a difference between reading and grinding. Reading 277 pages of mostly noise to find five pages of signal is not analysis. It is a tax on my time, and time is the one input I never get back.

Think back to the engineer and the $1,200 part. Musk didn’t refuse to understand the part. He refused to overpay for it. He broke it down to the raw metal and rebuilt the real cost himself. That is exactly the move here. I’m not refusing to understand the prospectus. I’m refusing to overpay for it in hours. So I broke it down to its raw inputs and let a machine do the reading that a machine should do.

I clipped the entire prospectus into the same Obsidian vault that already held Damodaran’s piece and the bull and bear cases from X. Then I pointed Claude Code at the vault Or its no-code twin, if a terminal isn’t your thing. More on that in a second.

This is where the actual work begins. And this is the part that works on any document you ever have to analyse, not just this one.

THE WORKFLOW: REFUSE THE PRICE, REBUILD THE NUMBER

A quick word on setup, because I’ve covered it in full before and won’t repeat all of it here. Pick your path, then come back.

If you’re using Claude Code, it’s not a coding tool. It’s Claude with hands, the same Claude you talk to in the chat window, except this version can read the files on your computer. I run it inside Obsidian, pointed at the vault where I just clipped the prospectus. If you’ve never wired this up, I walked through the whole thing in how I set up Claude Code as my investment research analyst. Start there.

If a terminal isn’t your thing, use Claude Cowork instead. It’s the same agentic engine in a friendlier room, no commands, you just point it at a folder and ask in plain English. I’ve shown exactly how I use it for a full equity report before, in how I used Claude Cowork to write a full equity research report in 90 minutes. Same destination, gentler road.

Whichever path you’re on, the rest of this works identically. The prompts are the same. Only the room changes.

One caution for both. Give the tool a folder with a copy of the prospectus in it, not your entire vault, and never your whole drive. These agents can edit and delete files, so you fence them into a sandbox and let them work there.

Now the work. I don’t ask the prospectus open-ended questions. I run it through the same checklist I run on every idea, in the same order, because the order is the discipline. Skip ahead to valuation and you’ll talk yourself into anything.

STEP ZERO: THE FIVE-LINE TEST

Before any number, one human question. Can I explain this business in five lines to a ten-year-old? If I can’t, it’s outside my circle and no spreadsheet will save me.

Here’s the honest answer for SpaceX. It’s three businesses wearing one ticker. A rocket business that launches things into space cheaper than anyone else. A satellite internet business, Starlink, that beams broadband down from those rockets. And an AI business, xAI, that runs enormous computers. The first two I understand. The third is where it gets hard, and that’s worth admitting before I read a single figure.

That admission shapes everything. A company you can’t fully explain in five lines isn’t disqualified. But it raises the bar on what the filing has to prove.

THE FRAMEWORK, AND WHY THE PROMPT IS NOT THE POINT

Here is the part most people get wrong. They go looking for the perfect prompt to paste. But a prompt you copy only works on the company it was written for. What you actually want is the thing underneath the prompt, the question, and the thing underneath the question, the principle. Get those right and you can write the prompt yourself, for any company, forever.

So here is the whole framework in one breath. Every step is a principle. Each principle becomes a single question. Each question becomes one tightly worded prompt that forbids the machine from guessing. The order is fixed on purpose. Growth first, then cash, then debt, then the quality of the profit, and valuation last, always last, because if you let yourself peek at the price before you’ve checked whether the profit is real, you will reverse-engineer a reason to like it.

Five principles, in order:

1. Is revenue genuinely growing, across years, not one good quarter dressed up as a trend.

2. Can the business fund its own growth from its own cash, or does it survive on outside money.

3. Is the debt load sane, or has the company handed control to its lenders.

4. Does the reported profit show up as real cash, or only on paper. This is the one that catches frauds.

5. And only now, after all of that, is the price sane?

The pattern of every prompt never changes: tell the machine exactly where to read, forbid it from estimating, demand a page citation for every number, and make it say “not disclosed” out loud when the filing stays silent. A filing’s silences are information too.

STEP ONE: REAL SALES GROWTH

Read more

I have been in the equity market since 2018. I still make mistakes every single day.

I do not say that the way people say it when they want you to think they are humble. I say it because it is the truest thing I know about this business. I make a mistake, I sit with it, I rethink it, I learn the lesson, and then I make a slightly different mistake the next week. That is the whole job. Anyone who tells you the best investors stop making mistakes has not watched the best investors closely enough.

Look at Mohnish Pabrai and Micron. One of the sharpest value investors of his generation sold most of his Micron position in 2023, before one of the great runs in semiconductors. That is not a small investor fumbling. That is a master of the craft, and the market still handed him a regret he will carry for years.

Because that is what the market is. The market is a place of regrets.

You regret buying the stock. You regret selling the stock. You regret not buying it. You regret not selling it. You regret the opportunity you took and the one you let walk past. There is no version of this game where you escape the pile. Everyone has a pile. The best investors do not have no regrets. They just work, every single day, to make the pile a little smaller than it was last year.

I used to write this newsletter about tools. The hot thing that month. Sometimes it was Claude, sometimes Perplexity, sometimes Chatgpt, sometimes something else. And somewhere along the way I realised I was doing my readers a quiet disservice. Because a tool you learn this month is stale next month. A tool does not make you a better investor. The only thing that makes you a better investor is becoming one, slowly, over years, by learning from people a few mistakes ahead of you.

So I changed what this newsletter is. It is not about tools anymore. It is about the regrets I am trying not to repeat, and the systems I build so I do not. I will never be a perfect investor. Not in twenty years. What I can do is show you my mistakes early enough that some of them never have to become yours.

Most of what I build, I give away here, free, because the point was never to gate the work. It was to make more investors better at this. If you want every skill, every workflow, and the downloadable guides as they ship, there is a paid tier that keeps this whole thing running, and I am grateful to everyone who is on it. But this edition, like the work itself, is yours.

Subscribe now

This edition is one of those mistakes.

It is about the part of a company almost no equity investor reads. The part that has a legal claim on the cash flows and the assets before you ever see a rupee or a dollar. The debt stack sitting directly above your shares.

I have held companies whose equity story I knew cold and whose debt I had never truly read. I knew the growth narrative, the margins, the management. And I had no real idea which notes were secured against the actual assets, which covenant flipped if the rating changed, or how many new shares a quiet convertible settlement was about to print on top of mine. I found those things out the way most equity investors find them out. Afterwards.

That is a regret. And it is one of the most common and most expensive regrets in all of equity investing, because almost nobody does the work to avoid it. Not because they are lazy. Because the work is genuinely brutal. You are hunting through a 10-Q, cross-referencing three footnotes to find one put date, squinting at conversion language buried thirty pages deep, trying to assemble a picture the company has deliberately scattered across the document.

So I built something so I would never have to skip that work again. And so you do not have to either.

I am going to show you the exact case, the exact filing, and the exact build. And by the end you will not just have a skill. You will see the stack that has been sitting above your equity this whole time.

What Sits Above You

Everyone has a view on Carnival.

The cruise line that almost drowned in debt when the world stopped sailing. The one that loaded up on borrowings just to survive 2020 and 2021, and has spent every quarter since trying to dig back out. If you follow the stock, you know this story. The recovery. The deleveraging. The slow climb back.

This is what most of us look at when we follow a company like this.

Now let me ask you the question I could not answer for years on companies I actually owned.

As an equity holder in Carnival, what exactly sits above you?

Not “the debt.” That single light blue liabilities bar is hiding the only thing that matters here. Because debt is not one thing. It is a stack. And every layer in that stack has a legal claim on Carnival’s cash and Carnival’s ships that comes ahead of your shares.

I pulled Carnival’s most recent quarterly filing and read the debt note properly. Here is what that one bar is actually made of.

At the top, with the first claim, sits roughly three billion dollars of secured debt, backed by collateral and guaranteed by the operating subsidiaries. These get paid first, and they have assets pledged behind them.

Below that, the bulk of it. Around twenty one billion dollars of unsecured debt that is still guaranteed by the subsidiaries. Notes running out to 2033. Euro notes. Export credit facilities stretching all the way to 2037. No collateral, but a guarantee from the companies that actually operate the ships.

And below that, the weakest tier of the debt, a little over a billion dollars that is unsecured and carries no subsidiary guarantee at all. Structurally the most fragile borrowing Carnival has. And it is still ahead of you.

Three separate layers. Roughly twenty six billion dollars of claims. All of it stacked above the equity. All of it gets satisfied before a single dollar reaches a shareholder.

And here is the part that should bother you most. This stack does not sit still. In the most recent quarter alone, one set of convertible notes that had been sitting in that pile simply disappeared from the debt and reappeared as equity. Converted. New shares, printed on top of yours, while everyone was busy watching the price chart recover.

None of this is hidden. It is all in the filing. Every number I just gave you is disclosed, sitting in a debt note, available to anyone willing to read it.

That is exactly the problem.

It is disclosed, but it is scattered. The principal is in one table. The maturity is in another column. The guarantee structure is in the row headers. The conversion that reshaped your ownership is buried in the equity statement three pages away. To see the stack the way I just laid it out, you have to hunt through the document, cross reference footnotes, and assemble by hand a picture the company has deliberately spread across the entire filing.

So almost nobody does it. Not because they are lazy. Because it is genuinely brutal, and slow, and the price chart is right there looking like enough.

That is the regret I kept making. Knowing the equity story cold, and never reading the stack that sat above it.

This is how I make sure I never skip that work again. On Carnival, or on any company I own.

The Build, In Four Moves

I am not going to read that filing by hand. Not anymore.

Not because I cannot. Because I have done it enough times to know how it ends. I find the debt table. I start pulling tranches into a spreadsheet. Somewhere around the third Euro note I lose my place. I go hunting for a maturity date and end up reading a footnote about something else. An hour later I have half a table, I am not sure I caught everything, and I still have not touched the conversion language sitting three pages away in the equity statement. The work is real. It is just the kind of work that falls apart the moment my attention does.

So I wrote down the method once, in a form that runs the same way every time, and never gets bored on the third Euro note.

That written-down method is what I mean by a skill.

Let me be precise about this, because it is the heart of the whole edition. A skill is not a prompt I paste and lose. It is a small instruction file that holds my methodology permanently. The difference matters more than it sounds. A prompt you re-type from scratch each time forgets things. You forget to ask about make-whole provisions on one company, about the fundamental-change put on another, about the events that quietly moved the balances on a third. A skill does not forget. Every nuance I care about is written into it once, so every company I point it at gets the same thorough treatment, in the same shape, with the same discipline.

And here is the question I know you are asking, because it is the right one. If I build this looking at Carnival, how can it possibly work on Walmart, or Tesla, or an aircraft manufacturer, when their debt looks nothing alike?

Because the skill does not encode Carnival’s debt. It encodes how debt is disclosed.

This is the part worth slowing down on. Every company that files with the SEC, regardless of industry, discloses its debt the same way, because the accounting rules force it to. A debt footnote. Each instrument. Its principal. Its maturity. Its coupon. Whether it is secured or unsecured. Whether it is guaranteed. Where it ranks. The conversion, call, and put terms. The events that changed the balance over the period. Tesla’s convertibles, Walmart’s plain senior notes, an airline’s debt secured against its aircraft, Carnival’s three tiers stacked above the equity. The instruments are wildly different. The disclosure skeleton is identical.

So the skill never assumes what the debt is. It only knows the questions to ask of any filing, and it asks them every time. That is what makes one skill work across thousands of companies. Not luck. Structure.

There is one honest limit, and I will say it plainly because it is what makes the output trustworthy. The skill is only as complete as the filing. If a company buries a key term in an indenture that the 10-Q only references, the skill cannot invent it. But it will tell you that the term is missing and point you to where to go find it. It does not paper over the gap. It marks it.

The build itself is small. Four moves. Most of it is the model working while I drink my protein shake.

Before the moves, one thing about why this is worth your time even if you never look at a single bond.

What I am about to walk through is not really a debt lesson. It is a pattern. A mistake I kept making, turned into a small tool that stops me from making it again. The mistake here happens to be ignoring the debt stack. But the four moves are the same whether the skill you build checks a company’s debt, runs a macro lens, or forces a pre-mortem before you size a position. Learn the four moves once, and you can build a tool around any mistake you are tired of repeating. The debt skill is just my example. Yours will be different.

Share

Move one. Define the job.

Before I touch a filing, I write one sentence. Who the skill is for, what situation it runs in, what it produces. Mine:

An equity or credit investor, looking at any company with meaningful debt, who needs every instrument, term, and ranking pulled from the filing into one clean, sourced structure before they trust their view of the company.

That sentence is the filter for everything after it. It tells the skill to work at the instrument level, not the summary level. It says the company is a variable, never a constant. Skip this and you get a vague “summarize the debt” tool that hands back the same shallow view we are trying to escape.

Here is the prompt I actually used to pressure-test that sentence before building:

I want to build a reusable skill. Before we write anything, here is the
one-sentence job: [paste your sentence].

Ask me three sharp questions that would expose anything vague or missing
in that sentence. Do not write the skill yet.

It will create an iteration 1 of the Debt structure skill on the questions answered by you.

Move two. Teach Claude the structure of a good output.

This is the move people skip, and it is the one that matters most. I do not write the skill from my own head. I have spent couple of years half-knowing what a debt waterfall needs, and half-knowing is exactly how things get missed. So I hand Claude a real filing and have it work out, with me, the full structure of what a complete extraction must capture, before any skill exists.

This is where the hard-won detail comes from: every field a debt instrument has, every event that moves a balance, every footnote nuance, every gap that needs flagging. I am not asking Claude to remember debt. I am asking it to lay out the skeleton that any company’s debt is disclosed within.

The prompt:

Read the attached SEC filing's debt note. Do not summarize it.

Lay out the complete STRUCTURE a credit analyst would need to extract a
debt waterfall from ANY company's filing, not just this one. Organize it as:

1. The fields every individual debt instrument has (principal, maturity,
   coupon, secured/unsecured, guarantee, ranking, conversion/call/put terms,
   status).
2. The kinds of EVENTS that change a debt balance over a period (issuance,
   repayment, conversion, redemption, repurchase, amendment).
3. The qualitative footnote terms that matter for risk (collateral,
   covenants, make-whole, fundamental-change puts, special interest).
4. The categories of information that are often MISSING from a filing and
   must be pulled from other documents.

For each, note where in this filing it appeared, so I can see the structure
is real. This becomes the backbone of a reusable skill.

What comes back is the backbone: the fields, the events, the nuances, the gaps. The domain, laid out. Now the skill has something true to be built on.

Move three. Write the skill.

Now, and only now, I have Claude write the instruction file, using the structure from Move 2 and the discipline rules I insist on: read only the filing, cite every figure, flag what is missing, stay at the instrument level, never assume the company.

The prompt:

Using only the structure we just laid out, write a SKILL.md for a reusable
skill named debt-waterfall.

The skill must:
- Take any company's SEC filing as input and produce four outputs: a debt
  master table ordered by ranking, an events ledger, qualitative notes, and
  an open-items list.
- Read ONLY the filing given. Never use outside knowledge about the company.
- Put "NOT DISCLOSED" wherever the filing is silent, and record it in open
  items. Never fill a gap from memory.
- Cite the page and section for every figure and every legal term.
- Stay strictly at extraction. No valuation, no recommendation, no opinion.
- Work on any filer in any industry, because it encodes how debt is
  DISCLOSED, not any one company's debt.

Format as a standard SKILL.md with frontmatter, role, the four outputs,
reading instructions, output discipline, and what it does not do.

What comes back is the skill. I read it once, tighten a few lines, and lock it. Here it is in full, free to copy:

full debt-waterfall.md inline here

Read the rules in it. The one that governs everything: if a number is not in the filing, the skill does not supply it. That single rule is why I trust this output in a way I would never trust a raw AI answer.

Move four. Install it, run it, and carry the result into Excel.

Read more

During COVID, I screened the market every single morning.

I had no idea what I was doing.

I’d read something the night before, some ratio, some filter, some “this is what the great investors look at,” and by 9am I’d run the entire market through it. The next week I’d read something else, and the screen would change completely. The parameters I trusted back then make me wince now.

But it felt like work. It felt like being a real investor. Screen, shortlist, open the annual report of whatever came out on top, repeat. I was busy from open to close, so I told myself I was getting somewhere.

I wasn’t. Months of this, and I couldn’t point to one idea worth holding that the screen had actually handed me.

And here’s the part that took me years to admit. It wasn’t just me being a beginner.

When I started working in the markets professionally, everyone around me ran screens too. The junior analyst next to me. The PM. The most senior man on the floor, twenty years in, had his own filters he swore by. Different parameters, same ritual. Everyone was screening. Almost nobody was finding anything that way.

A couple of years back, I stopped.

Not because I found a better screen. Because I finally noticed where my good ideas were actually coming from, and it was never the screen. It was people. The investors I read. The ones I got on calls with, seasoned HNIs, operators who had been through three cycles, sometimes a junior analyst with one sharp observation. A writeup on Substack. A thread I couldn’t stop thinking about. Someone handing me not a ticker, but the question they were chasing.

That shift is the single best decision I have made as an investor. And it is exactly how I ran into the analysis I am about to show you.

I didn’t screen for it. I came across it: a video breaking down what the big four are doing with their AI spending, and it was the best take on the sector I'd seen in a long time. It stopped me cold.

Then it left me with one question.

Could I rebuild that myself? To that standard?

With Claude Code.

I know how that sounds. Most people in this business will tell you it can't be done, that serious research takes a data terminal, a team of analysts, and a budget none of us have. I'm not going to argue with them. I'm going to show you the whole thing instead, every prompt, every number, and the places I caught it getting things wrong. Don't take my word for any of it. Judge it for yourself.

This is the story of what happened when I tried.

What Claude Code Pulled First, and Why It Looks Like Madness

Here's the entire first pass in one clip, start to finish. I open the terminal inside Obsidian, paste in the prompt, and Claude Code goes out to the SEC filings on its own, pulls capex, operating cash flow and free cash flow for all four companies, and writes it all back into the vault. I handed it nothing. If you watch one thing in this piece, watch this.

I didn’t ask Claude Code for an opinion. I already had the video’s. I gave it the idea and pointed it at the source documents instead. The 10-Ks, the 10-Qs, the earnings releases. The filings, not a summary of them.

I never left Obsidian to do it. Claude Code runs right inside the vault, in a terminal you add from the community plugins in about a minute.

My whole research library sits on the left, Claude Code works on the right, and everything it finds gets saved back into the same place. Nothing scatters across tabs and tools.

Prompt 1: ok so I just watched this breakdown on the big tech names, Microsoft and Amazon mostly, and the whole thing is about how they're spending insane money on AI data centers right now and the market is basically punishing them for it. I want to check if that fear is actually real or if everyone's just overreacting, like is the market right to be scared here or not.
I haven't given you any filings, so you'll have to go get them yourself. pull the latest 10-Ks, the most recent 10-Qs, the earnings call transcripts, whatever you need, straight from the source, EDGAR or their investor relations pages. don't use some random article or summary, I want it from the actual filings.
first thing I want to understand is the spending. how much capex did Amazon and Microsoft actually spend last year, what are they guiding to for this year, and put Google and Meta next to them too so I can see the full picture across all four. give me the year over year jump in plain numbers.
then the part that actually matters, the cash. for Amazon show me operating cash flow vs capex and what's left as free cash flow, and what that does to the free cash flow margin. same for Microsoft. if there are analyst estimates anywhere for where free cash flow lands this year, pull those too, I want to see how bad people think it gets.
and don't just dump numbers on me. based on what's in the filings, tell me straight, is the market right to be this scared? and save whatever you pull into the vault so we don't lose it, I'm going to keep building on this.

The first thing it did was rebuild the spending picture from the filings. And the first number it handed back was the one that scares everyone.

Amazon spent about 77.7 billion dollars on net capex in 2024. In 2025 that jumped to 128.3 billion. A 65 percent increase in a single year. And on the latest earnings call, Jassy guided to roughly 200 billion for 2026. Microsoft ran the same play. Its capital spending went from 44.5 billion in fiscal 2024 to 64.6 billion in fiscal 2025, up 45 percent, and it has only accelerated since: the December 2025 quarter alone was roughly 37.5 billion, so even that fiscal 2025 number understates where the run rate sits today.

Then Claude Code pulled the cash, which is where it actually stings.

Amazon generated 139.5 billion in operating cash flow last year. It spent 128.3 billion of it on capex. What was left as free cash flow was 11.2 billion. The year before, that number was 38.2 billion. Free cash flow fell 71 percent in twelve months, and the share of operating cash flow Amazon got to keep dropped from 33 percent to 8 percent.

Read that cold, off the screen, and it looks like the biggest companies on earth setting fire to their own cash at the same time. That is exactly how the market read it. It is why the selloff happened.

And here is the trap. If you stopped reading the filings right here, you would sell too. Most people did.

But before I trusted a single one of those numbers, there was one step left. The one I never skip.

Rule One: I Never Trust the First Pass

The first output looked finished. Clean tables, sources listed, a confident take at the bottom. That polish is exactly when you should get suspicious.

So before I believed a single number, I made Claude Code audit its own work. Not re-read its notes, that’s worthless, it just nods along with itself. I sent it back to the actual filings, line by line, told it to recompute every number it had derived, and to assume it had gotten at least one thing wrong.

Prompt 2: ok before I trust any of this, go back and check your own work, and assume you got at least one thing wrong, because you usually do. don't just re-read your notes, that's pointless. go back to the actual filings you pulled from and re-confirm every raw number against the source, capex, operating cash flow, free cash flow, revenue, all of it, and tell me the exact line or page you got each one from.
then recompute everything you derived yourself, the year over year percentages, the free cash flow margins, the capex as a percent of revenue. don't trust the numbers you wrote down last time, redo the math from the base figures and see if it still matches.
and cross check against one outside source too, like the company's own release against an aggregator, and if they disagree tell me where and why, because sometimes it's just a definition thing like net vs gross capex.
then give me a clean list of what was right, what was wrong, and what you changed, and update the file so the corrected version is the one that stays in the vault.

It had. Two things, in fact. It had Microsoft’s fiscal 2025 revenue about 20 billion dollars too low, which had quietly thrown off the capex-intensity math built on top of it. And it had overstated Meta’s capex, because the first pass folded Meta’s finance-lease payments into the raw spending line.

These sound small. They are not. A wrong revenue number poisons every ratio sitting above it. And here is the one that would have actually embarrassed me: three of these four companies define free cash flow differently. Amazon nets out proceeds from asset sales. Microsoft and Alphabet don’t. Meta subtracts finance-lease payments on top of capex. Compare Meta’s free cash flow to Amazon’s without catching that, and you are quietly comparing two different things, then drawing a conclusion from a gap that doesn’t exist.

The first pass of any AI research will have errors. That is not the failure. The failure is trusting it. The discipline is the second pass, and a second pass that forces the model back to the source instead of letting it grade its own homework. That single step is the whole distance between research you can put your name on and the confident-sounding slop filling everyone’s feed.

Claude Code didn’t stop here. The next thing it pulled is what flipped the whole picture.

The Money Didn’t Burn. It Turned Into Assets.

So I went back and asked the other side of the question. If the cash is leaving, where is it going? I pointed Claude Code at the same filings, plus Amazon’s 2025 shareholder letter, to pull the part of the story the free cash flow panic leaves out.

Prompt 3: ok so the last numbers make these companies look like they're torching cash, but I don't buy that read and I want to test the other side. the whole point is capex isn't normal spending. when they buy data centers and chips they end up owning assets that earn for years, it's not gone like salaries or marketing.
first, go back into the same filings and show me the depreciation side. how much depreciation are they running now, how fast is it growing, and explain in plain terms how the capex they spend today turns into a depreciation hit over the next few years. I want to know what the real drag on reported profit actually is, because I think that's the only thing genuinely getting worse, and it's temporary next to how long these assets earn.
second, and this one matters most, pull from Amazon's own 2025 shareholder letter. Jassy talks about how the market reacted to AWS spending back around 2014 and how it played out. I want that parallel in their own words, the part about the spend and the long term and that "straight line was a lie" idea. keep any quotes short.
save it next to the capex file, and like before, don't trust your first pass, I'll verify after.

Here is what the screen doesn’t show you. When a company pays salaries or runs ads, the money is gone the moment it’s spent. When Amazon spends 128 billion dollars on data centers and chips, it does not vanish. It becomes 128 billion dollars of assets sitting on the balance sheet, earning for years.

The reason it looks like a wound is timing. Capex hits the cash flow statement all at once, today. But it reaches earnings slowly, spread out as depreciation over the life of the asset, four to six years for servers, far longer for the buildings. So the cash looks terrible now, while the earnings hit is still building.

And that hit is real. This is not a story where nothing is wrong. Depreciation is climbing fast. Microsoft’s depreciation and amortization jumped 53 percent in a single year, to 34 billion dollars. Amazon's total depreciation and amortization came to 65.8 billion for full-year 2025, up about 25 percent in a single year, and it keeps climbing as more of that capex lands on the income statement. Claude Code pulled the balance sheets to show why: Microsoft’s property and equipment grew by 69 billion dollars in one year, and Amazon’s AWS assets nearly doubled. All of that is only now starting to depreciate.

But here is the part that flips it. The worst of the drag is ahead, not behind, because capex is still growing, so depreciation keeps rising for two to three more years before it peaks. The fear that crushed these stocks was about the cash. The real cost is the depreciation. And depreciation is temporary next to how long the assets earn.

Amazon has run this exact movie before, and Jassy says so himself. In the 2025 letter he goes back to 2014, when AWS was deep in its own spending and a senior leader inside the company openly questioned why they were even in the business. The market disliked the spend then too. His point is that long-term bets never travel in a straight line, they zig up, stall, and double back. He borrows a phrase for it: “straight line was a lie.” Then the comparison that lands it. Three years after AWS launched, its revenue run rate was only in the tens of millions. Three years into the AI wave, Amazon says AWS’s AI run rate is already past 15 billion dollars, roughly 260 times larger at the same stage.

In 2006, as Amazon poured money into the infrastructure that became AWS, BusinessWeek ran a skeptical Bezos on its cover under the headline "Amazon's Risky Bet." A Piper Jaffray analyst caught Wall Street's mood: "I have yet to see how these investments are producing any profit." The spending was called a distraction. That distraction is now AWS. Bezos keeps the cover framed.

That is the whole reframe. The cash that looks burned is being converted into what Amazon believes is its next AWS-scale pillar, and the only genuine cost, depreciation, fades as the assets earn.

One honest caveat, because it’s the hinge the whole thing turns on. This holds only if the AI revenue actually shows up. If demand disappoints, that rising depreciation stops being temporary and becomes a years-long drag with nothing behind it. So the question is no longer “why are they burning cash.” It is 'when do these assets start paying, and is the demand really there.' But before I let any of that into the thesis, I ran it through the same check as the first set of numbers.

Read more

For the past couple of months, I have been writing about Claude Code for investment research, and the reaction split almost perfectly in two.

The people who actually tried it became serious fans. They got past the black terminal and realised the funny part: you do not need to be a developer to use it. You can talk to it in plain English, point it at your annual reports, your transcripts, your messy folder of consultancy PDFs, and it starts working like a research analyst sitting inside your laptop.

Most people never crossed that first barrier. They saw the terminal and quietly checked out, and I understand exactly why. Finance professionals did not spend years learning valuation, earnings power and management behaviour just to feel stupid in front of a blinking cursor. Humanity has invented many forms of unnecessary pain, and command-line anxiety is apparently one of them.

So today I want to step one layer simpler. Not away from Claude Code, because I still believe the CLI tools are where serious AI research eventually goes. But before most investors get there, they need something easier to hold.

That brings us to Claude Skills.

I started building Skills very early, right after Anthropic shipped them, and for three or four months I was obsessed. I built finance skills, broke them, rebuilt them, debugged prompts for hours, and tried to turn every repeatable piece of my research into a reusable system. It felt like real progress. The outputs looked the part, clean and structured, the kind of thing you screenshot and post.

And then something happened that I did not notice for weeks. I stopped opening them.

Nothing broke. The equity research skill still ran, the 13F screen still produced a tidy output, I just drifted away from my own work without realising it, until one day I sat with the uncomfortable question of why. The honest answer was that not one of those skills had ever changed a decision I actually made. The output was beautiful and my conviction sat exactly where it would have been without any of it. I had built a very sophisticated way of feeling like a more serious analyst.

Claude Skills are useful. But they are not the end goal.

Most people are approaching them completely wrong. They download somebody else’s investment Skill because it looks sophisticated from the outside, maybe it says Buffett-style investing or 13F framework or equity research copilot, they use it for a week, the outputs look impressive, and then they quietly stop, because the Skill never actually helped them build conviction.

That is the real problem, and it runs deeper than tooling. Investment research is deeply personal. Two successful investors can agree on eighty percent of the principles and still work completely differently, because the way they extract information, the way they connect dots, the questions they keep asking and the things they choose to ignore are never the same. You cannot inherit that from a zip file.

So this post is not a download-this-and-you-are-done tutorial. I am going to build a real investment research Skill in front of you, show you the structure, the output and the mistakes I made, but the bigger point is the one underneath it. A skill will not make you a better investor. What it can do, if you build it yourself, is force your own thinking into the open, the questions you always ask, the ones you forget, and the gaps you have been quietly walking around for years.

So here is the question I would sit with if I were you. That skill you installed last week, the equity research copilot, the Buffett framework, the 13F tracker someone handed you in a zip file. Be honest. Has it changed a single decision you have made? Or does it just make you feel like the kind of analyst who uses tools like that?

What Actually Happens After You Build Investment Skills For 90 Days

I assumed the hard part of all this would be the prompt engineering. It was not.

The hard part was how honest it forced me to be about my own research. One day I would care deeply about management commentary. The next I would lose three hours inside capital allocation. Some weeks I obsessed over industry structure, other weeks I skipped it entirely and spent all my attention on whether earnings could actually last. I had always told myself I had a process. The moment I tried to turn it into reusable instructions, I ran straight into how much of it was not a process at all.

That is what building a skill does, and it does it brutally. You see which questions you ask every single time, which ones you forget, which parts of your workflow are real signal, and which parts are just there to make you feel sophisticated while adding almost nothing to conviction.

That was the point where skills finally became useful to me. Not because they automated research, they cannot, but because building them forced me to be honest about which parts of my process were actually repeatable and worth keeping. And once I saw that, I stopped trying to build one giant equity research brain and started building small skills around the specific moments where consistency actually matters. That worked far better.

You can see this the moment two people build the same skill. Hand the same earnings-review brief to a margins-obsessed analyst and a management-quality analyst and you get two different tools. One builds it to interrogate the cost line and operating leverage. The other builds it to track what management promised last quarter against what they delivered this one. Same name on the folder, completely different instructions inside.

So before I write a single line of a SKILL.md file, before I touch folder structure or decide whether the output is a memo or a checklist, I force myself through four questions.

The Four Questions I Answer Before I Build Anything

This sounds basic, but it is where almost every skill goes wrong. Equity research is too broad. Investment research is too broad. Even analyse a company is too broad. None of those are workflows. They are ambitions, and a skill built on an ambition will try to do everything and end up doing nothing you actually trust.

A useful skill sits inside one specific moment of your process. Reviewing an earnings call the morning after results. Comparing management commentary across four quarters. Checking whether capital allocation has quietly drifted over five years. Turning a 10-K into a list of open questions. Reading expert call notes and pulling out the things that could break the thesis. Once the moment is that narrow, the skill no longer has to pretend it can replace you. It only has to do one repeated thing well, every time.

This is also where most downloaded skills fall apart. They try to be a full analyst, because a full analyst sounds impressive in a product description. But in real research the leverage almost never comes from replacing the whole process. It comes from making one repeated step sharper and more consistent than you manage on your own.

So for this post I am not going to build a complete equity research analyst. It would look impressive and it would be useless by the third run. I am going to build something deliberately narrow. A thesis-breaker.

The job of this skill is simple. You give it your thesis, along with whatever you have collected, the notes, the filings, the transcripts, and it goes looking for what could break it. It does not tell you to buy or sell. It does not pretend to know the future. It forces your own assumptions into the open and asks the questions you are most tempted to skip, the ones where the answer might cost you the position. Whether the earnings assumptions are too clean. Whether management commentary has quietly changed. Whether the valuation already prices in the good news. Whether the entire thesis is leaning on one variable behaving perfectly.

That is where AI actually helps a serious investor. Not by becoming Buffett in a zip file. By becoming a consistent sparring partner around one specific moment that matters.

Once the moment is clear, the next decision is not what the skill should do. It is what it should never do. This matters more in investment research than almost anywhere else, because AI gets confident far too quickly. Ask it to sound intelligent and it will. Ask it to analyse a company and it will hand you a clean, polished answer even when the evidence underneath is thin, stale or missing. A beautiful memo built on weak evidence is still weak research. It just looks more respectable while it misleads you.

So before I write the workflow, I write the boundaries. For the thesis-breaker they are simple. It does not give buy, sell or hold calls. It does not pretend missing data does not matter. It does not invent numbers, peer comparisons, management quotes or valuation assumptions. And it does not attack the thesis just to sound contrarian, because that failure is as useless as flattery. A thesis-breaker that lists every possible risk in the world looks analytical and says nothing.

A good one separates the risks that actually threaten the thesis from background noise. If the thesis depends on margin expansion, it goes after what breaks margins. If it depends on operating leverage, it tests whether revenue growth is really enough to carry the fixed cost base. If it depends on management, it compares what they said before with what they are doing now. The guardrails are not compliance decoration. They are what stops the skill from quietly turning into a confident generalist. The single rule underneath all of them: the output should make me think harder, never make the decision for me.

I do not want the input to be clean, because real research never is. Sometimes you have a half-written thesis. Sometimes a page of notes from a call. Sometimes a valuation assumption sitting in your head that you have never actually written down. A useful skill has to take that messy material, work out what it has and what it is missing, and still separate the core thesis from the noise.

So the input is flexible but not lazy. At a minimum it asks for the company, the current thesis, whatever source material exists, and the specific part of the thesis you want pressure-tested. If something is missing, the skill does not paper over it. It says what it has, what it does not, and which parts of the analysis will be weaker as a result. That one choice matters more than it looks. A research skill should never reward messy thinking with a clean answer. Give it two lines of thesis and it should tell you the read is preliminary. Give it no numbers and it should not invent any.

I do not want a long memo just because long memos feel serious. The output has to be short enough to read in the middle of an actual research session and sharp enough to change what I do next. For the thesis-breaker, it returns five things: my thesis in plain English, the assumptions hiding inside it, the strongest objections to it, the evidence I am still missing, and the questions I should answer before I add to the position.

That is the full design. A narrow moment, clear boundaries, realistic inputs, a useful output. Now I can stop describing the skill and actually build it.

Building the Thesis-Breaker

Now that the design is clear, I can turn it into something real. And the first decision is the one most people get wrong.

For the first version, I am keeping this deliberately simple. One folder. One SKILL.md file. That is it.

This matters because people often make Skills look more complicated than they need to be. They create multiple files, nested folders, examples, checklists, and instruction documents before the Skill has even proved that it is useful. That may look serious, but it is not always efficient.

The file tells Claude when to use the skill, what workflow to follow, what to produce, and what it must never do. If the same instructions are going to run every time, they can simply live in that one file. I would only split it later, and only if the skill earns it. If the workflow grows too large, the methodology can move into its own file. If the guardrails get detailed, they can move into another. If I start adding worked examples across sectors, those get a folder. But none of that on day one. A multi-file skill is not automatically more advanced. It is only better if the extra files genuinely reduce confusion, and on the first run they never do. They just hide the fact that you have not tested anything yet.

The goal is not the most impressive skill folder. It is the simplest version that actually works, fast enough that you can point it at a real thesis today.

Here is the prompt I would give Claude or Claude Code to build the first version.

The rest of this issue is for members. What follows is the part I can’t fake for you: the skill run live on a real Alphabet thesis, where it broke, and what it told me to build next. I opened Alpha with AI to paid members three days ago. For the next stretch, until 8 June, the first year is 25% off, my thank-you to the people who showed up before there was ever a wall.

Claim 25% off →

Read more