The Bluff Filter is this kind of check, on one page. Take it with you →
// On this page
This is not another stock-research prompt list. It is a decision checkpoint for after the research and before the order, when a polished thesis can make one weak assumption difficult to see.
Five questions, in order. First the bet itself, last the dimension missing from the frame. Each uses AI as a critic, organiser or calculator; none delegates the buy decision. These assume you’ve already done the background research. If you haven’t, that’s a different job: the free AI tools worth using for it are covered separately. Come back here once you’ve got a view and the current sources behind it.
The five questions: (1) What am I actually betting on? (2) What’s the strongest case against buying this today? (3) What does the market already believe? (4) What would have to change for me to be wrong? (5) What am I not asking?
What am I actually betting on?
The story behind a buy (strong brand, good management, big market) is not yet a testable bet. I want one specific claim that has to remain true for the expected return to materialise. When the answer comes back as a vague paragraph, I push again until it collapses to one testable sentence, then check that sentence against my own evidence. It is a candidate statement of the bet, not something the model gets to declare true. Check it is about the right company while you’re at it: Gemini once audited an entirely different business than the one I pointed it at, fluently, and never flagged the swap.
What’s the strongest case against buying this today?
For a stock decision, there’s a difference between a bad company and a bad entry. A company can fit a long-term thesis while a confirmed event makes the timing unattractive. This question asks for the timing case rather than another generic risk register. It only works if the calendar and claims are current: an invented earnings date or competitor launch is not a bear case. When I run it, I’m not looking to be talked out of the buy; I’m looking for the near-term fact or plausible scenario I’d kick myself for ignoring.
What does the market already believe?
Reverse-engineering a stock’s implied valuation is an arithmetic question, not a story prompt. A model cannot infer what the market believes from a ticker alone. Give it the current price, diluted share count, cash, debt, forecast period and valuation method, then make it show which growth and margin assumptions reconcile the model. I read the result to compare those explicit assumptions with my own. A gap is a question to investigate: it could mean the market sees something I missed, or that my valuation model is wrong.
What would have to change for me to be wrong?
A thesis you can’t falsify isn’t a thesis. This is not “what could go wrong” but “which reported metric or observable event would show that the core assumption is failing?” The relevant window may be one quarter or several years; forcing every thesis into the next three months creates false precision. I use the model to propose candidates and locate the reporting source. I choose and write down the threshold before buying, because otherwise I know I can move the goalposts later.
What am I not asking?
The last question asks for a dimension missing from the frame. The other four test claims already inside the thesis; this one asks a model to compare the thesis with a fixed checklist and propose an omission.
That can be useful, but it is not independent judgement. A model can echo the framing, agree with the user or invent a precise-sounding consequence. Give it a short summary, ask it to check competitive, financial, timing, regulatory and structural dimensions, then make it separate a sourced omission from a speculative question.
In the original May 2026 write-up, I recorded running this on a META thesis framed around AI-driven ad optimisation and capital discipline. Gemini raised the regulatory dimension. The useful part was the omitted category: the DMA requires consent before a gatekeeper combines or cross-uses personal data across designated services. The saved response went further, speculating about structural decoupling and asking for an unsupported percentage. That was a question to investigate, not a sourced finding.
// What Gemini said
Dimension flagged: Regulatory
If antitrust mandates or privacy legislation (such as the EU’s Digital Markets Act) formally decouple Meta’s ability to synthesise data across Instagram, WhatsApp, and Facebook, by what specific percentage does the “proprietary data” moat degrade, and can the AI engine maintain its targeting superiority without that cross-platform signal?
How would a structural decoupling of Meta’s data silos change your assessment of a 40% operating margin floor?
That transcript is preserved exactly as the article recorded it. No raw capture survives elsewhere in the repository, so this recheck does not independently verify or re-sign the wording, the META thesis or the model session.
Where these fall short
The questions structure the reasoning. They do not replace it. Access to prices, filings, options chains and market data depends on the exact chat, connected app and account; a connected route can retrieve data that a standard chat cannot. Either way, provenance and timing still matter.
Investor.gov warns against relying solely on AI-generated information for investment decisions because it may be inaccurate, incomplete, misleading or outdated. Verify each decision input against the filing, company calendar or broker before it touches the trade. Use the answers to test your own thinking, not as instructions to buy. Ask the model to make the call itself and the confident narrative can hide which numbers it actually checked, which is the whole reason these questions keep you, not the AI, in the chair.
The decisions these questions protect against are the ones you’d already half-made before you sat down: the buy you were looking for permission to make, not advice on whether to make it. Take the time the evidence needs before the money is on the table. The Bluff Filter is the compact source-and-risk check; the broader four-stage framework these questions draw from is the Prompt Stack.
The pair to this one runs in the opposite direction. These five questions guard the buy. There is a matching prompt for the sell decision. It does the same job in reverse, forcing the exit trigger to be written down before the conviction has somewhere to escape to.
The short version
What worked: The sequence gives the model bounded jobs: reduce the thesis, build a sourced timing challenge, show valuation arithmetic, propose observable falsifiers and check a fixed list of missing dimensions.
What didn’t: An unsourced prompt cannot establish current events, market data, an embedded valuation or a valid sell threshold. Supply the inputs, demand provenance and make the decision yourself.
Time required: However long it takes to assemble and verify the current inputs. The prompts are short; the evidence check is the work.
The original write-up says I ran these questions against three buys before writing them down: NFLX, GOOGL and TXRH. The supporting trade ledger is private, and this recheck does not verify or re-sign those positions. The sell side of the same discipline follows the same pattern.
Ben tests how far you can trust the main AI assistants, and publishes exactly where they get things wrong. Every post here is a first-hand test with the receipts, including the times a tool simply wasn’t worth the trust. About Ben →
The site tests how far you can trust the main AI assistants, on real decisions. Start with the Prompt Stack for the four-stage framework, free and ungated, or the Bluff Filter for the paste-ready version with a real before and after.