The AI prompt I run before every sell decision

I’ve held META while its capex outlook kept climbing, capex being the money a company spends building things, in META’s case the data centres behind its AI push. Meta opened its 2026 range at $115–135bn in January, then the April Q1 release raised it to $125–145bn, well clear of the $120bn peak I’d assumed when I bought. For the first time, I ran an AI prompt before the sell decision, a structured thesis audit rather than a gut check. My recorded read from that run: the original bounded-capex thesis was gone. What I’m doing about that (hold, sell some, or exit) is my decision, and a separate question from whether the prompt did its job.

That separation matters. The point of the prompt is not to maximise returns on any single holding. It’s to make sure I’m not making sell decisions on price when the real question is whether my reason to own still holds. The output is a defensible thesis read, not a trade.

In my 20 May 2026 search of competing sell-prompt pages, each one asked AI to recommend an exit or invent selling signals. That starts in the wrong place. The first question is whether the reason you bought still holds. Only after that audit is there a basis for discussing an action.

What most sell-prompt content gets wrong

The 20 May scan split into two patterns. Some pages generated exit signals such as valuation peaks or earnings disappointments. Others buried a vague “act as a financial adviser and tell me whether to hold” among roundup prompts. Both began at the recommendation instead of the thesis.

The standard mistake is starting with a recommendation. A general “should I sell?” question can be answered with general qualifications about risk tolerance, time horizon and market conditions. All true, all too loose to test against the reason you bought.

The harder part of the sell decision isn’t getting a recommendation. It’s checking whether the thesis you built when you bought still holds. The prompt structures that task: the input is your original thesis, the output is whether it survives what has happened since.

The standard sell-prompt asks AI to recommend. This one asks AI to audit.

The prompt has four sections (SCOPE, FILTER, RISK, VERDICT) following the Prompt Stack structure. The SCOPE makes the model audit your thesis before discussing an action. The FILTER forces you to type your original thesis before reading any output. The RISK names the rationalisation you are most likely to make. Only then does the VERDICT ask for a coarse, thesis-anchored action rather than a price reaction.

The AI thesis-audit prompt

I run this prompt every time I notice a position straining. The FILTER section forces me to commit the original thesis in writing before reading any output, and that act is most of the work. It follows the Prompt Stack: SCOPE, FILTER, RISK, VERDICT. The placeholders in square brackets are yours to fill in. The FILTER section is where the work happens: if you can’t state your original thesis in three sentences, the problem is the thesis, not the prompt.

The SCOPE keeps the model on the thesis before it reaches an action. The FILTER forces you to type your original thesis before reading anything the model returns. That act is the behavioural intervention. The VERDICT then closes with thesis-anchored action, not price-anchored action. Price-anchored: “it’s down 12%, cut your loss.” Thesis-anchored: “the free-cash-flow thesis still holds, but capex guidance is a legitimate flag: sell half.” (Free cash flow: the cash a business has left after running itself and building for the future. It’s what funds the dividend, the buyback, or the next bet.)

What it caught on META

The 1 May 2026 META run used this FILTER, recorded after the post-earnings drop.

• Position: Long META, held since late 2024 / early 2025, average price paid roughly $500, current price around $610.
• Original thesis: META is investing through an AI infrastructure cycle. Ad revenue growth is durable. The free-cash-flow trajectory makes the valuation defensible if capex is bounded. My assumption when I bought was annual capex would peak somewhere around $120bn.
• What has happened since: FY25 came in at $72.2bn. Management then opened FY26 guidance at $115–135bn in January and raised it at Q1 results in April to $125–145bn. The floor of the latest band is above the $120bn peak I’d assumed, and its top end is over 20% higher.

Question 1 made the standard case for AI investment: ad revenue growth is intact, the spend gets absorbed as revenue grows, and the squeeze on free cash flow is a timing problem, not a permanent one.

Question 2 was the one that did the work. My notes from the 1 May run paraphrase it this way: repeated increases break the bounded-capex assumption, not just strain it; the original thesis treated $120bn as a peak, while the new range starts above it. That was sharper than the sentence I would have written for myself. It made the inference rather than restating the input, the difference between “capex is rising” and “the assumption your thesis rested on is gone.”

The RISK call landed harder still. My notes record the rationalisation as telling myself the AI investment thesis justifies the capex, which is also management’s case. That’s the sentence you read and wince at, because you’ve already started constructing the justification before you finished the FILTER.

What I do with this read is a separate question from whether the prompt did its job. My recorded conclusion from the run was that the bounded-capex assumption had gone. The VERDICT can offer hold, sell some or exit, but it cannot take responsibility for that decision. Mine to weigh.

The prompt's job ends at the thesis read. The trade is yours.

Three weeks on, I ran a simpler version through Claude (Opus 4.7, with live web search, 22 May 2026), sell-some-vs-hold on the same META holding, with the same capex-raise context. The prompt’s discipline still holds. Claude reframed the bounded-capex break in sharper language than my original Q2 paraphrase: “the floor of 2026 guidance now sits above the ceiling you assumed.”

The case for selling some, in Claude’s read: the part of the spending driven by rising memory-chip prices is “the worse kind of capex increase, because you’re paying more for the same capacity rather than buying more capacity. That’s a margin signal, not a growth signal.” The honest caveat in the same response: “selling a compounder into capex fear is historically how people have given up Meta’s best runs.” Different framing of the question; the same fork.

Where the prompt falls short

Six caveats worth being honest about up front. I learned most of these the hard way running this on real positions; worth reading before you run it on yours.

It requires the original thesis in writing. If you cannot state in three sentences why you bought, the prompt returns a generic audit of a stock you’ve described, not an audit of your thesis. Most retail holders don’t have their entry thesis written down. The real value is forcing them to write it.

The honest read depends on what you put in the FILTER. If you summarise the negative evidence accurately, the model produces a useful synthesis. If you frame it charitably, the model will too.

AI cannot audit what you haven't disclosed.

Its action is coarse, not a sizing decision. The VERDICT gives you “hold, sell some, or exit”, not “sell exactly 30%.” How much depends on how big the holding is, the price you paid, and the rest of your portfolio. Treat that line as an output to inspect, not an instruction to execute.

It is most useful before price pressure, not under it. Running it after the stock is down 15% is better than not running it, but the RISK stage is harder to take seriously when you’re already down. The discipline is to run it at a calm moment, when you notice the thesis is straining.

The prompt takes your stated thesis as given. Its SCOPE tells the model to work from what you supply. If your original thesis was wrong from the start, this prompt may audit it on its own terms without testing the original error. It audits consistency, not correctness. The 5 questions to ask AI before buying any stock covers the other side of the discipline.

It’s a thesis check, not a portfolio tool. Some sell decisions aren’t thesis-driven at all: selling some because a position got too big, raising cash, or cases where the thesis is fine but the money would do more work elsewhere. The prompt doesn’t help with any of those. Confusing the two produces worse decisions than running no prompt at all.

The sell decision is not the hard part. Recalling what you were actually betting on, and being honest about whether that has changed, is the hard part. The prompt makes you do the recalling before the price tells you what to think. It won’t place the trade for you, which is just as well. That part should still cost you a little sleep.

The Prompt Stack covers the underlying methodology, free and ungated. The buy side of the same discipline is at 5 questions to ask AI before buying any stock. Every published failure from running these prompts is collected on lessons.

The private trade ledger remains private; this article carries only the bounded details needed to explain the method.