A fluent answer still needs checking

One answer. One thing to verify.

You do not need to investigate every sentence. Ask for the exact source of the claim that matters most, open it, and check whether it actually supports the answer.

The full 30-second method

The Bluff Filter.

The five-stage check to keep beside your next AI conversation. It comes with Confidently Wrong.

The Bluff Filter asks an AI to name its sources and separate claims from guesses. Use it to organise what needs checking.

1. USE: what will you do with the answer?
2. SCOPE: which sources, dates and limits apply?
3. FILTER: what is known, inferred or guessed?
4. RISK: what could make the answer wrong?
5. VERDICT: one useful action, with honest confidence.

Download the complete instructions · Read the PDF

Seven checks, one at a time.

Choose the failure you suspect. Copy one instruction. Inspect the dated receipt behind it. Open only the check you need.

None makes the model perfect. Each makes one failure easier to spot.

Want the whole set? The Bluff Filter combines the same checks into one instruction for the top of a chat.

Behind the seven: nine ways AI gets it wrong, the dated evidence and six assistants compared.

The Prompt Stack, for a bigger decision.

A way to organise an AI-assisted answer: set its scope, inspect its claims, look for counterevidence and choose a next step. The prompt is a starting point. Check the facts that matter before relying on the answer.

The Prompt Stack uses four stages: SCOPE sets boundaries; FILTER asks for evidence and labels; RISK asks what could change the conclusion; VERDICT asks for a next step. None verifies the answer by itself. Check decisive claims against the original sources. A second AI can reveal a disagreement worth investigating, but agreement is not proof.

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 next step, with its assumptions and limits.

Method note · 25 September 2026. Earlier wording overstated what these prompts could establish. Claim labels, model confidence and agreement between assistants are not verification. The examples and current downloads are the existing method; a revised checking workflow is under review.

Copy the prompt

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: two questions, before you trust a word

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. 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. A material error is a reason to stop relying on the unchecked answer. It does not prove that every other sentence is false. Recheck the claims that carry your decision. Passing these two questions only establishes relevance; it does not certify the answer.

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.

Those checks give you a starting point. The four stages below ask the AI to expose its assumptions and gaps. You still need to inspect the evidence behind the claims that matter to your decision.

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 fold below.

One example, both ways: weighing up a holiday deal

The basic ask leaves the source boundary open. The more specific ask tells it to use the supplied page and identify missing information. Check that it followed the instruction before relying on the answer.

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

Telling an AI to be sceptical or to act as an expert is not a fact-check. A more specific request is: 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. That request sets a boundary; the answer still needs checking against the page.

02 · FILTER: make it label every claim.

An AI answer can mix evidence, inference and guesses in the same confident tone. Ask it to tag claims as [sourced] with a destination, [inferred] or [guessed]. These are the model's own labels. A sourced tag can still point to the wrong page or misread a real one. Open the evidence for the claims your decision depends on; the labels only help organise that check.

On that holiday deal, "breakfast included [sourced: the deal page]" gives you a place to check. "Breakfast included [guessed]" names a gap. Read the deal page before relying on either: the tag does not establish what is included.

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.

04 · VERDICT: one next step, with its limits.

Ask for a practical next step and the assumption it depends on. That step may be to find a missing fact, rather than make a decision. Low, medium and high are self-reported confidence labels, not measured probabilities. Asking again can expose an inconsistency, but a repeated answer can repeat the same mistake.

If a decisive fact is missing, the useful verdict is what to check next.

That is the whole method. Aim it at anything you are being asked to trust on faith: copy all four stages as one prompt, edit the bits that are yours, and make it show its working.

How do you run all four stages in one prompt?

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.

Copy it and paste in the answer or decision you are checking, with the material it should use. Treat the output as a list of claims and questions to inspect. You can also run the stages separately and check each output before continuing.

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 next step with one sentence of why. State which assumption in List B it depends on most. If a decisive fact is missing, make checking it the next step. Describe what remains uncertain; do not substitute a confidence label for evidence.

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.)

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.

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.

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.

Give one practical research step with one sentence of rationale. State which assumption in List B it depends on most, and which evidence remains unchecked. Do not force a trading decision when a decisive fact is missing.
Keep the next step concise and state its limits.

When it really matters: cross-check it.

A second assistant can suggest another interpretation or expose a disagreement worth investigating. The Scoreboard records how assistants answered the same questions on particular dates. Two assistants can also repeat the same error or rely on the same source. Their agreement does not verify a claim.

Use the comparison to decide what to inspect, then check the decisive facts against the original material. Questions that need professional judgement still need that judgement; another AI answer cannot supply it.

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.

The Source Ladder: how close is the source?

The Source Ladder is a five-rung way to check a source: rank it by how close it sits to the original. Rung 1 is the primary source itself, the law or the filing or the dataset. Rung 5 is pure speculation. When an answer matters, push yourself, and the AI, up the ladder toward rung 1. It is a filter, not a law. Here is the ladder, and a real case of it catching a slip.

You will not always reach rung 1, and you do not need to for a trivial question. The point is simple: the more a claim can cost you if it is wrong, the higher up the ladder you should insist on standing before you trust it.

The ladder, catching one.

Worked example · Gemini (Pro) · 7 July 2026

The claim: the maximum UK fine for using a handheld phone while driving a lorry or bus is £2,500.

Gemini got the number right. That fine is published on gov.uk, a rung-1 fact. Beside the figure it displayed a Police.uk source label, but the captured label exposed no destination URL. The correct gov.uk guide appeared only as a detached closing source. An opaque label has no rung because there is no page to inspect.

The danger is that an official-looking label can feel like a checked citation. The ladder catches it with one question: what exact page is this? If the answer exposes no page, stop there. If it does, climb to the most direct source and find the claim on it. Police.uk now carries the same figure, but the historical destination is unknown.

The full run: AI cites the wrong source (8 July 2026), where three of five assistants cited a real page that did not back their claim.

Cited rung 3, claimed rung 1.

That is how a catch or a logged failure records a slip like Gemini's in one line. The claim was a rung-1 fact. The citation behind it was a rung-3 source dressed up as proof of it. The number sitting where the source should be tells you the exact size of the gap: a genuine primary source would read "claimed rung 1, cited rung 1". When the two numbers diverge, that space between them is where a confident answer goes wrong.

A page can straddle rungs, so the shorthand is used only when the fit is clean. It is a way to see the gap, not a label to force onto every miss.

The ladder is the reader's version of the source-tier grading the site already runs on the Scoreboard. For why we grade the way we do, see How we grade.