The One Prompt That Stops Your AI From Sounding Sure When It Isn’t

By the end of this guide, you’ll have one saved prompt you can paste in immediately after any AI answer you’re about to act on.

It might be a price, a recommendation, a summary of a contract, or an assessment of a customer. The prompt takes the answer you already have and forces it to show you how sure it actually is — and what would prove it wrong.

Keep a real decision in mind while you read: something you’d normally just trust and move on from.

The problem this solves

Ask an AI tool “Should I raise prices on this product?” or “Does this contract let the customer cancel early?” and it will answer in the same confident, complete-sounding tone whether it actually knows or is filling in a plausible-sounding guess.

There’s no built-in signal for “I’m quite sure” versus “This is my best guess based on limited information.” You have to ask for that signal explicitly. It doesn’t show up on its own.

This matters most exactly when you’re least able to check it yourself: a legal clause you’re not trained to read, a market comparison you don’t have time to verify, or a customer’s intent you’re guessing at from a short email.

The dangerous combination

The more difficult an answer is for you to verify, the more persuasive confident wording can feel. Tone is not evidence.

Build It: The Confidence Check

Ask your question as normal and get an answer. Then, in the exact same chat — don’t open a new one — paste this below the answer and press enter:

Confidence checkPaste after the answer

Go back through your answer above. For each claim you made,
give me:
1. A confidence level: high / medium / low
2. What specific information you actually had to support it
   (not "generally" or "typically" — the actual source: what I
   told you, or general knowledge with no specifics)
3. One piece of counter-evidence or a scenario that would make
   this claim wrong
4. If confidence is medium or low: what I'd need to find out
   before acting on this

That’s the whole prompt. Use it after any answer that matters — not every chat message, just the ones you’re about to act on.

What comes back

Claim: “$450 is competitive for this service in your area.”
1. Confidence: Low
2. What I had: General knowledge of typical pricing for this type of service nationally. I do not have data on your specific area or your local competitors.
3. Counter-evidence: Prices can vary significantly between regions because of cost of living and local competition. Without local data, this figure could be too high or too low.
4. What to find out: Check prices from two or three competitors in your actual service area before setting this price.

If what you get back doesn’t look like this — no confidence level, no real answer to “What did you actually have?”, just a restated version of the original answer — ask again:

Follow-upUse only if needed

Answer all four numbered points specifically, using the format I gave you.

What to do with the result

Act

High confidence

If the counter-evidence is genuinely weak or doesn’t apply to your situation, you can act on the answer.

Check first

Medium or low confidence

If the counter-evidence sounds plausible, don’t act yet. Complete the “what to find out” step first.

When in doubt between the two, treat it as the second case. A wrong price, a missed contract clause, or a wrong read on a customer costs more than the extra five minutes.

Why each part earns its place

A confidence level, not just an answer — this is the signal missing from the original response. Without asking, you get an answer with nothing attached to it.
What information it actually had — this reveals the difference between “I gave it the exact contract clause” and “It’s guessing based on how these contracts usually work.” Those levels of reliability read identically without this question.
Counter-evidence, required every time — this does the real work. Asked to defend its answer, an AI will usually find supporting arguments. Asked what would prove it wrong, it has to look for the weak point instead of reinforcing its first answer.
What to find out when confidence is low — this turns “I’m not sure” into an actual next action instead of a shrug you have to solve yourself.

Two quick examples

Pricing

Is $450 competitive?

You receive a confident yes. Run the check. If confidence is low and the answer is based on general knowledge rather than your area or competitors, that yes was a guess dressed up as an answer.

Worth knowing before you set the price.

Contract

Can the customer cancel?

You paste in the clause. High confidence tied to the exact sentence, with no applicable counter-evidence, is much stronger.

If another sentence elsewhere contradicts it, the check may catch the problem before you act.

When the check itself isn’t enough

If the confidence check keeps returning “high” on things that later turn out to be wrong, the issue usually isn’t the check. It’s that the AI never had the actual source material to begin with.

A confidence check can only be as good as what was in the conversation. If you asked a general question without pasting in the actual contract, notes, or data, the model has nothing concrete to rate itself against. “High confidence” may only mean “This sounds right based on how these things usually go.”

Paste in the actual source material first. Run the check second.

What You Now Have

One saved prompt that turns a confident-sounding AI answer into four visible signals: how sure it is, what it’s actually based on, what could make it wrong, and what to check before you rely on it.

You’re not getting a better first answer. You’re getting a way to tell, before you act, whether the answer you already have deserves your trust or your double-check.