Skip to content
// The Method

The Prompt Stack.

A method for getting answers out of AI you can act on, instead of the confident, fluent, sometimes-wrong ones it hands you by default. One quick check on any answer, then four questions you run in order. It works on a holiday deal or a doctor's letter; I push it hardest where the answer has to be right.

// The method, in order: four stages, top to bottom
  1. SCOPE

    Telling it to be sceptical changes the tone, not the answer. Fence it in: say what it may use, and make it admit when something is past what it can verify.

  2. FILTER

    Make it label every claim: sourced, inferred or guessed. The guessed ones are where you look first.

  3. RISK

    Name the one thing that would prove the answer wrong, something you can check.

  4. VERDICT

    One answer, with a confidence level. No fence-sitting.

// On this page
// Before the four stages

How do you tell if an AI answer is true before you act on it?

AI sounds exactly as sure when it's wrong as when it's right. A made-up figure arrives in the same calm, helpful voice as a correct one, so tone tells you nothing, and tone is what most of us are quietly judging on. So before any of the four stages, I run one quick check on the answer in front of me. It takes under a minute, and it's aimed at one specific trap: an answer that's confident, fluent, and quietly about something other than your problem.

// The check is two questions: ask them before you trust a word
  1. 1. Name the exact thing. Make it tell you precisely what it's looking at (the specific page, document or item) before it summarises anything. "Which one is this, exactly?"
  2. 2. Hand me one fact I can check in under a minute. One date, one number, one name I can hold against the real thing myself.

Check One, Bin the Lot. Fail either question and the whole answer goes. Not the wrong line, all of it. The good-looking paragraphs aren't a consolation prize. They're the part that nearly fooled you. Passing both questions doesn't make the answer right; it means it's about the right thing, and worth running the four stages on.

Say you've pasted in your phone contract and asked for a plain summary. Before you read a word of it, make the AI tell you what it's looking at: which provider, which plan, the monthly price printed at the top. If it says £29 when the page in front of you says £39, that's someone else's contract being described in a very reassuring voice. Then make it hand you one fact you can check fast. A made-up number gives itself away the second you look. A right answer about the wrong thing never does: every fact correct, neatly laid out, and about a contract you've never seen in your life.

// The quick check: paste this before you act on an AI answer Before you answer anything else: name exactly what you're looking at, the specific document or page in front of you. Then give me one fact about it I can check myself in under a minute. If you're not sure what you're looking at, say so before you go any further.

That clears the first trap. The four stages below are for the answer once you trust it's about the right thing. They stop an answer that's technically correct but missing what matters from talking you into a decision.

// The detail

What are the four stages of the Prompt Stack?

One question gets you one cheerful answer. Four questions, asked in order, each clear the ground for the next. Here's each one with an everyday example; the exact prompts I run on real money are in the proof layer below.

// One example, both ways: weighing up a holiday deal
The basic ask
Is this holiday deal a good buy?
The better ask
Use only the deal page I have pasted below, nothing you remember about the resort or current prices. If anything I need is not on the page, or is the sort of thing that changes, say so rather than guess. Is this holiday deal a good buy?

The basic ask lets it pad the answer with half-remembered prices and a glowing write-up of a hotel it has never seen. The fenced-in ask sticks to the page in front of it and admits the bits it can't check. Same four-stage method below, on real money.

01 · SCOPE
Fence it in, and let it say "I don't know".

The obvious first move is to tell it to be sceptical, or hand it a job title. Neither buys you much. "Be critical" changes the tone of the answer, not how right it is, and "act as a top analyst" is a costume, not a fact-check. What does cut the error is narrower: tell it exactly what it's allowed to draw on, and give it permission to stop. "Answer only from the document I pasted." "Stick to the current UK rules; if you're not sure they're current, say so." An "I'm not sure, that's past what I can verify" is worth more than a confident guess that reads just as well.

Weighing up a holiday deal, I fence it to the deal page itself: don't price the resort from memory, don't fill in the bits that aren't shown. And I tell it plainly that "the page doesn't say" is an allowed answer. It stops inventing a glowing review of a hotel it has never seen and starts telling me which of my questions the page actually answers.

The exact prompt I run on a real position is in the proof layer below ↓

02 · FILTER
Make it label every claim.

Most AI answers mix what's on the page with what the model invented on top, same tone, same confidence. Asking it to "be honest about what it's unsure of" doesn't fix that; a model has no reliable sense of its own gaps. What works is making the answer label itself. I tell it to tag each claim as it goes: [sourced] with where it came from, [inferred] if it's a reasonable read between the lines, [guessed] if it's filling a hole. Now you're reading a map of what to trust instead of one smooth paragraph, and the guessed tags tell you exactly where to look first.

On that holiday deal, "breakfast included [sourced: the deal page]" is one thing; "breakfast included [guessed]" is quite another. Same five words, and the tag is the difference between a fact you can act on and a hole you need to plug yourself.

The filter run on a real company, with the invented lines struck out, is in the proof layer below ↓

03 · RISK
Name the one thing that would prove it wrong.

"What are the risks?" gets you a checklist, and a checklist isn't the point. The question that matters is the one almost nobody asks: what exact thing would prove this wrong, a tripwire, not a vague worry. A risk you can't see in advance is background noise. A risk with a specific signal is something you can act on.

On the holiday deal: "if the total at checkout comes out over £900, the headline price was a lie." Now you're not weighing a foggy feeling. You've got one number to watch, and the deal either trips it or it doesn't.

The exact prompt I run on a real position is in the proof layer below ↓

04 · VERDICT
One answer, with a confidence level.

Nobody standing at the checkout with their card out has ever been helped by a balanced six-paragraph essay on both sides. The last step stops the hedging: one answer, and how sure it is, low, medium or high. A confidence label you can hold it to later is worth more than a summary it can hide behind. The honest upgrade, when it matters, is to ask the same thing again a different way and see if the answer holds. A verdict that survives a re-ask is one you can lean on.

Book it or don't, in one line, with a confidence level. One verdict you can decide on, not a fog you have to referee.

The exact prompt I run on a real position is in the proof layer below ↓

// Try it now

That is the whole method. Run it on a question of your own: build a ready-to-paste prompt in a few clicks, or copy the combined version below and edit it yourself.

// Run it on anything

All four stages as a single copy-paste prompt, for any answer you need to trust or any decision you are weighing. The exact versions I run on real money are further down.

// TL;DR

How do you run all four stages in one prompt?

Copy it, paste in the answer or decision you are checking, run. Works on anything with a right answer: a contract clause, a letter from a doctor, a deal, a position. Or use the per-stage prompts further down for tighter control between steps.

// The combined prompt: ~200 words, copy & run
All four stages in one prompt.
// The Prompt Stack: combined You are checking one answer, or working through one decision, in four stages, in order. Do not combine stages or skip ahead. STAGE 1: SCOPE Work only from what I give you below, plus sources you can name. Do not fill gaps from memory. If something is the sort that changes (a current price, a live date, a figure that moves), or a question is past what you can verify, say so plainly instead of guessing. STAGE 2: FILTER Before giving any answer, produce two lists, and label each claim: LIST A: Observable facts, verifiable from what I gave you or a named source. Tag each one [sourced: where it came from]. LIST B: Assumptions and inferences, anything you are reading between the lines, extrapolating, or filling in yourself. Tag each one [inferred] or [guessed]. If uncertain which list something belongs on, put it in List B. Do not proceed to Stage 3 until both lists exist. STAGE 3: RISK Based on List A and List B only, identify: (a) What would make this answer or decision wrong (b) The single thing that most has to be true for it to hold (c) One specific signal I could check that would tell me it is going wrong STAGE 4: VERDICT Give one practical answer or action, and a confidence level (Low / Medium / High), with one sentence of why. State which assumption in List B it depends on most. No hedged summary. One answer. What I'm checking: [paste the answer, or describe the decision, with any context] (When it really matters: run this same prompt through a second AI from a different lab, and dig wherever the two answers disagree.)

Running each stage separately lets you review and correct the output between steps. Worth doing when real money is involved. Per-stage prompts below ↓

Or skip the copy-and-edit: build your own version with a few clicks →, then paste it into whatever AI you use.

// When it really matters

When it really matters: cross-check it.

For anything that actually matters, run the same question through a second assistant from a different lab. Where the two disagree is exactly where to dig. It's the whole idea behind the Scoreboard: different models fail in different places, so two of them agreeing narrows the odds, and two of them disagreeing points straight at the thing worth verifying.

It costs a minute and one extra paste. On a holiday deal it's overkill. On a medical question, a contract, or a real position, it's the cheapest insurance going. Asking the same model twice rarely catches its own mistake; a second model, built differently, often does.

// Where I push it hardest: real money

The same four stages, on a real investment position.

Where a confident wrong answer does real damage. Here are the exact prompts I run on a live position, with a worked example, and the posts where each one met a real decision.

01 · SCOPE
Fence in the sources before the question.

Ask "what do you think of this stock?" and the model reaches for whatever it half-remembers: a price that's months stale, a metric from a filing it never saw, a confident figure with no source. The first stage closes that door. Tell it exactly what it may use, the numbers you've pasted plus sources it can name, and tell it to flag, not fill, anything it can't stand behind. A stale price quietly stated as today's is the failure this stage is built to stop.

// Stage 1: Scope Work only from the figures and text I give you below, plus published sources you can name (filings, audited accounts, regulatory announcements, market data). Do not fill gaps from memory. If a figure is the sort that moves (a current price, a live date, a market figure) and you can't verify it from what I've given you, say so rather than state a number. If a question is past what you can confirm, say "I can't verify that" instead of guessing.

A grand job title ("act as a top analyst") and a "be sceptical" instruction both feel like they should help, and neither does much: they change the register, not the accuracy. Constraining the source does. If you want an adversarial read on top, add it, but it's the fence around what the model may draw on that earns its place here.

02 · FILTER
Separate facts from filler.

Most AI output on a company looks like analysis and is actually narrative: verifiable facts stitched together with plausible-sounding extrapolations, in the same tone, with the same confidence. The filter step makes it sort the claims into two lists and label each one, so a guess can't sit in the answer wearing a fact's clothes.

// Stage 2: Filter Before forming any view on this position, produce two lists and tag every line: LIST A: Observable facts, verifiable from published sources (filings, announcements, audited accounts, market data). Tag each one [sourced: name the source]. Do not include anything that requires interpretation to establish. LIST B: Assumptions and inferences, anything depending on management guidance, analyst projections, extrapolation, or interpretation. Tag each one [inferred] or [guessed]. If uncertain which list something belongs on, put it in List B. Do not skip this step. Do not move to analysis until both lists exist.

After the output: scan List A and move anything that's actually an inference. Models smuggle assumptions into List A constantly, and a missing or vague [sourced: …] tag is the tell. Phrases like "the company is well-positioned to…" are not facts. Once the lists are clean and every line is tagged, everything that follows is built on something auditable.

03 · RISK
Make the downside explicit.

"What are the risks?" gives you a checklist, and a checklist isn't what this stage is for. The third question, what you'd actually see if it was going wrong, is the one that matters. A risk you can't observe in advance is background worry. A risk with a specific signal is something you can act on.

// Stage 3: Risk Based on List A and List B only, identify: (a) What could go wrong on a 12-month view (b) What would have to be true for the entire investment case to be wrong (c) What observable signal in the next 90 days would tell you it's going wrong, something specific enough to act on If you can't give a specific answer to (c), say so.

If the AI can't give you a specific observable signal for (c), it doesn't really have a view. It's pattern-matching off the consensus. That's a useful signal in itself.

04 · VERDICT
One action. With a confidence level.

By this point there's something worth reading. The verdict step is short on purpose: one practical action, a stated confidence level, and a sentence on why. An AI forced to commit to "low confidence" is more useful than one allowed to hide behind a balanced summary.

// Stage 4: Verdict Give one practical action (buy / add / hold / reduce / avoid) and a confidence level (Low / Medium / High) with one sentence of rationale. State which assumption in List B the verdict depends on most. No hedged summary. One action.

The confidence label builds up into a track record over time. Was it right when it sounded sure? What happened when it wasn't? That calibration is most of the value.

// FILTER · stage 02

What does the FILTER stage actually remove from an AI analysis?

A typical AI response on a company mixes verifiable facts and invented narrative in the same breath: same tone, same confidence. Stage 02 pulls them apart. Here's what it cut from a META Q4 2024 analysis.

// Meta Q4 2024 · filter step applied
// List A · kept
  • Revenue $48.4bn, up 21% year-on-year
  • Operating income $23.4bn, margin 48%
  • Family of Apps MAU: 3.35bn
  • Reality Labs operating loss: $5.0bn
  • Q1 2025 guidance: $39.5–41.5bn
// List B · cut
  • "continued dominance in digital advertising"
  • "exemplary management execution"
  • "unassailable position in social media"
  • "Reality Labs showing encouraging progress"
  • "signals strong management confidence"
The strikethrough items are what a typical AI response adds on top of the source material. The filter step removes them before you act on anything.
// Prompts in practice

Which posts apply the Prompt Stack to a real decision?

Each of these posts is the Prompt Stack applied to a real decision, with the actual model output, the finding, and what I did with it.

For the full per-tool failure catalogue (every AI fabrication caught across these tests, with the prompt, the output, and the screenshot) see The Lessons.

// The paste-ready version

Want the four-line version plus a real before and after? The Bluff Filter is one short instruction set you paste into your AI once, so from then on it scopes where each answer came from and flags every guess.

Get the Bluff Filter →
Ben Dixon
// The author

I test how far you can trust the main AI assistants, and I publish exactly where they get things wrong. Everything here is a first-hand test with the receipts, including the times a tool simply wasn’t worth the trust. More about me →

// Open methodology · The Prompt Stack is also published as an open repository at github.com/CtrlCursor/prompt-stack under CC-BY 4.0. Share, adapt, credit.