The Truth Prompt: Stop AI From Confidently Making Things Up

AI does not lie the way a person lies. There is no intent. It is a prediction engine, and a sentence that is confidently wrong reads just as smoothly as one that is confidently right. That is the whole problem.

The Truth Prompt is one block of instructions you paste into your AI's settings once. From then on, every new chat is forced to separate fact from guess, score its own confidence, and tell you out loud when it does not know.

This guide is the full version: the prompt itself, setup for every major tool, why it works, and the follow-ups I use to keep AI honest in the moment.

The Truth Prompt

Copy the whole thing. Do not shorten it. Paste it into your AI's custom instructions and save.

THE TRUTH PROMPT (COPY ALL OF IT)

Adopt the role of a rigorous, truth-seeking reasoning assistant.

For every complex or factual request, follow this process:

1. DECOMPOSE
Break the request into smaller questions or claims that can be evaluated independently.

2. DISTINGUISH
Clearly separate:
- Verified facts
- Reasonable inferences
- Assumptions
- Opinions
- Unknown or missing information

3. SOLVE
Address each part carefully. Do not invent facts, sources, quotes, statistics, links, or details to fill gaps.

4. VERIFY
Before answering, check:
- Logical consistency
- Factual accuracy
- Whether the answer fully addresses the request
- Whether important context is missing
- Whether bias or unsupported assumptions may be affecting the answer

5. CALIBRATE CONFIDENCE
Assign a confidence score from 0.0 to 1.0 based on the quality of the available evidence, not on how persuasive the answer sounds.

6. RETRY WHEN NEEDED
If confidence is below 0.8:
- Identify the weakest claims
- Reconsider the reasoning
- Revise the answer
- Ask for clarification or state what information is needed if the uncertainty cannot be resolved

7. BE HONEST ABOUT UNCERTAINTY
Never present an assumption, prediction, or uncertain claim as a confirmed fact. If something cannot be verified, say so directly.

OUTPUT FORMAT
For every final response, use this format:

Clear Answer:
Provide the most accurate and useful answer possible.

Confidence Level:
Give an overall confidence score from 0.0 to 1.0 and briefly explain it.

Key Caveats:
List any assumptions, uncertainties, missing information, conflicting evidence, or facts that require verification.

Why AI Confidently Makes Things Up

A language model is a prediction engine. It looks at everything in front of it and predicts what text should come next. It is not looking things up in a database. It is not checking a source. It is producing the most plausible continuation of a sentence, and a sentence that is confidently wrong is just as plausible as a sentence that is confidently right.

Researchers call this a hallucination. It shows up most often as fake citations, invented statistics, made-up URLs, quotes nobody said, and function names that do not exist. The output is fluent. That is exactly what makes it dangerous.

The three failure modes you actually run into:

1. Gap filling. The model knows 80 percent of the answer and smooths over the missing 20 percent instead of flagging it. You never see the seam.
2. Confidence without evidence. The model's tone is the same whether it is certain or guessing. You read confidence and assume accuracy. They are unrelated.
3. Agreeing with you. Models are trained on human feedback, and humans rate agreeable answers higher. Push back on a correct answer and watch it fold. That is not the model changing its mind, that is it optimizing for your approval.

What The Prompt Actually Does

The Truth Prompt does not give the model new knowledge. It changes the model's process before it answers and forces it to show you its work afterward.

Think of it as the difference between asking someone a question in the hallway versus asking them to write you a memo. Same person, same knowledge. Wildly different rigor.

The seven steps each exist for a reason:

• Decompose splits the request so a wrong sub-claim cannot hide inside a right-sounding paragraph.
• Distinguish forces every piece to be labeled fact, inference, assumption, opinion, or unknown. Most bad AI answers blend all five.
• Solve bans gap filling. Naming the exact things it fabricates works better than "be accurate."
• Verify is the second pass that catches contradictions the first pass wrote in.
• Calibrate separates "I am sure" from "this reads well." Those are different things.
• Retry gives the model a chance to catch itself before you ever see the answer.
• Be honest gives explicit permission to not have an answer. Without it, silence feels like failure to the model.

The Clear Answer, Confidence Level, Key Caveats output format is half the value. The caveats section is the one people skip and it is the most useful part of the whole thing. It is where the model tells you what it was unsure about, what it assumed, and what you should go verify yourself.

Install It Everywhere

Do this once per tool and it applies to every conversation from then on.

Claude (web and desktop)

Click your name in the bottom left. Go to Settings. Find "Instructions for Claude" (also called personal preferences). Paste the full prompt and save. If you want it on one project only rather than everywhere, open a Project instead, click "Set project instructions," and paste it there.

ChatGPT

Click your profile in the bottom left. Settings, then Personalization, then Custom Instructions. Paste the prompt into the box labeled "How would you like ChatGPT to respond?" and save. If the full prompt does not fit the character limit, use the compressed version below.

Gemini

Open Settings, then Saved Info (or Personalization, depending on your version). Add the prompt as a saved instruction. Gemini's memory is less strict than Claude's instructions field, so re-paste the prompt at the top of any high-stakes chat as a backup.

Claude Code (or any coding agent)

Put it in your CLAUDE.md file at the root of your project. Every session reads that file automatically. For coding, add one extra line under the main prompt:

Additional rule for code:
Never invent library functions, API endpoints, flags, or config keys. If you are not certain a function exists, say so and tell me to check the docs. A wrong function name that looks right costs more time than saying "I do not know."

API or system prompt

Drop it in as the system message. Same text, no changes. If you are building anything that touches facts, numbers, or citations, this belongs in your system prompt permanently.

Short Versions For When You Need Them

The full prompt is the right one for custom instructions. But sometimes you need something that fits in a character limit, or you just want to spot-check one answer.

The compressed version:

Be a rigorous, truth-seeking assistant.

For factual requests: break the question into separate claims. Label each as verified fact, inference, assumption, or unknown. Never invent sources, stats, quotes, or links to fill gaps. Check your answer for logic, accuracy, and missing context before replying. If your confidence is under 0.8, revise it or tell me what you would need to know.

End every factual answer with:
Confidence: 0.0-1.0 (based on evidence, not on how good the answer sounds)
Caveats: assumptions, unknowns, things to verify

The one-liner:

Answer this, then give me a confidence score from 0.0 to 1.0 and list every assumption or unverified claim you made. Do not invent sources or numbers. If you do not know, say so.

The audit (run it on an answer you already got):

Re-read your last answer as a skeptical fact-checker who wants to find errors.

List every factual claim you made. For each one, mark it: VERIFIED, LIKELY, UNCERTAIN, or FABRICATED.

Then rewrite the answer keeping only what survives.

The audit is the highest-leverage prompt in this guide. Run it on any AI answer you are about to act on. You will be surprised how often something gets downgraded.

How To Read The Confidence Score

The score is useful. It is not a measurement. The model is not computing a probability, it is estimating one, using the same prediction machinery that produces everything else it says. That means the score is a signal, not a guarantee. A 0.9 is not a promise. It is the model saying "this feels well-supported to me."

So why bother? Because the relative movement is informative. When a model that normally hands you 0.9 suddenly hands you 0.55, something in that answer is shaky, and it just told you.

Rough guide to what the numbers should mean:

• 0.9 to 1.0. Well-established and widely documented. Proceed. Still verify anything with money, health, or legal consequences.
• 0.75 to 0.9. Solid but with assumptions baked in. Read the caveats carefully. That is where the soft spot is.
• 0.5 to 0.75. The model is reasoning, not recalling. Real uncertainty. Do not act on this without a second source.
• Below 0.5. It is guessing and admitting it. Treat as a starting point for your own research.

A high confidence score on a hallucinated fact is still possible. The prompt reduces this. It does not eliminate it. If an answer will cost you money, time, or credibility when wrong, verify it against a real source.

The follow-up that breaks a fake high score:

You gave that a high confidence score. Now argue against yourself.

What is the strongest case that your answer is wrong? What would I find if I checked your sources and they did not exist? Where is the weakest link in your reasoning?

Then re-score your confidence.

8 Follow-Ups That Force It To Stay Honest

Custom instructions set the baseline. These are what you type in the moment, when an answer feels a little too smooth.

1. The source check

For every source, statistic, and quote you just used: do you actually know this exists, or are you reconstructing what a plausible source would say? Be specific about which is which.

2. The anti-agreement check

I think you changed your answer because I pushed, not because I was right. Were you correct the first time? Tell me honestly.

3. The missing context check

What did you leave out of that answer that a real expert would have included? What would they say you got dangerously wrong?

4. The disconfirming evidence check

Give me the strongest evidence AGAINST what you just told me. Do not soften it.

5. The blank-slate check

Forget everything I told you I wanted. If you had no idea what answer I was hoping for, what would you say?

6. The verification path

Do not tell me the answer. Tell me exactly how I would verify it myself, and what I should search for. Then give the answer.

7. The knowledge boundary check

Is this something you actually have reliable training data on, or is it recent, niche, or obscure enough that you are extrapolating? Be specific about where your knowledge thins out.

8. The stakes check

I am going to act on this. Money and reputation are on the line. Given that, what do you want to change or caveat about your answer?

Number 2 is the one you will use most. Models fold under pressure. If you push back and the AI immediately agrees with you, that is a red flag, not a win.

The Mistakes That Turn The Prompt Off

The prompt works, but there are specific things people do that quietly cancel it out, and then they conclude it did not work.

• Pasting it into an old chat. Custom instructions only load at the start of a conversation. Start a new one.
• Trimming it down to "be accurate." The model already thinks it is being accurate. What works is the specific list of things it is banned from inventing (sources, quotes, statistics, links) and the required output format.
• Ignoring the caveats section. People read the Clear Answer, glance at the confidence number, and skip the caveats. The caveats are the payload.
• Treating 0.9 as verified. It is the model's self-assessment. For anything that matters, you still check.
• Leaving it on for creative work. If you are brainstorming or writing fiction, this prompt makes the AI stiff and hedgy. It will caveat a poem. Turn it off or scope it to a Project.
• Assuming it stops all hallucination. It does not. Nothing does. It makes hallucination less frequent and, more importantly, more visible.

The single biggest gain from this prompt is not fewer errors. It is that the errors now announce themselves.

When To Use It, When Not To

Rigor has a cost. Answers get longer, more hedged, and slower. That is a great trade for research and a terrible one for a quick idea.

Turn it on for research and fact-finding, anything with statistics or numbers, medical, legal, or financial questions, writing you will publish under your name, code you will ship, decisions with real money attached, and anything you will cite to someone else.

Turn it off for brainstorming and idea generation, creative writing and fiction, casual conversation, first drafts you will heavily rewrite, rough scratch code, and rewording or summarizing tasks where you already know the answer.

Pair it with search when you can. The prompt makes the model honest about what it does not know. Web search gives it a way to actually find out. Together they are far stronger than either alone.

Test It Yourself In Two Minutes

Do not take my word for it. Open a normal chat with the prompt off and ask something the model cannot possibly know. Something niche, recent, or fully invented. Watch it confidently answer. Then start a new chat with the prompt installed and ask the same thing. Watch it flag the question, refuse to invent details, and hand you a low confidence score with real caveats.

Try the same test with a topic in your own field, where you will instantly spot a fake answer. It is a lot more convincing when you catch it yourself.

The 5-Minute Action Plan

1. Paste the full prompt into Claude's Instructions for Claude and save.
2. Start a NEW chat.
3. Ask it a fake-study bait question and watch it refuse.
4. Save the audit prompt somewhere you can grab it.
5. From now on, read the Key Caveats section before you act on anything.

An AI that says "I do not know" is worth ten that sound sure.