Prompt That Stops AI Hallucinations (Claim Labelling)
Splits the answer into three layers: verifiable fact, inference and guess. The model learns to say "I don't know", and you see which sentences you can trust.
You have a habit that causes most of your errors: producing confident sentences on subjects you do not actually know. In this task we switch that habit off. SUBJECT: [SUBJECT] WHAT I AM USING THIS FOR: [PURPOSE] START EVERY SENTENCE WITH ONE OF THREE TAGS [VERIFIABLE] — the claim has an identifiable source; a date, number or name can be produced [INFERENCE] — reasoned from known material, but no source can be cited [NOT SURE] — no reliable grounding in your training data for this claim RULES 1. If you write a number, date, rate, statute name or person's name, put the source in brackets on the same line (institution + year). Never invent a source. If you do not have one, use [NOT SURE]. 2. Do not guess and keep going. Stop and write "this needs verification". 3. End the answer with three lists: - MUST VERIFY: claims that have to be checked before publication - WHERE TO CHECK: one primary source type per claim (official body, statute text, the manufacturer's own documentation) - RISKIEST SENTENCES: the three sentences that would do the most damage if wrong 4. Write in plain English. No inflated adjectives. Avoid empty certainty: "definitely", "everyone knows", "it is widely accepted". Process the subject now. One-paragraph summary first, then the tagged body.
This does not eliminate hallucination; it makes it visible. You still have to open the sources behind every [VERIFIABLE] line before you publish.
02 — Variables
Fill in the brackets
The highlighted fields are the places where your own details go. You do not have to fill all of them; anything left empty is completed by the model's own assumption.
[SUBJECT]
[PURPOSE]
[VERIFIABLE]
[INFERENCE]
[NOT SURE]
03 — How to use
How to use it, step by step
- 01
Start narrow. Not "tax" but "2026 UK income tax bands" — a narrow subject gives the model less room to invent.
- 02
After the first answer, copy out only the [VERIFIABLE] lines. Delete the other two categories from your draft.
- 03
Run the same prompt a second time with: "now attack only the [NOT SURE] lines from your first answer and show why each one cannot be supported." Those are the riskiest claims.
- 04
Never click a URL the model gave you and call it checked. Search the institution's own site for the number.
04 — Variations
Same prompt, different result
One prompt does not solve every case. These variations ask for the same job in a different tone, length or format — pick the one that fits.
Citation-forced
Same subject, one extra rule: write nothing you cannot source. Output verifiable information only. Then add a section called "What I cannot tell you" with three sub-headings you would normally skip — name the gaps instead of hiding them.
Default choice for legal, medical and financial text.
Red team
Read the text below like a prosecutor. Sort every sentence into: provable / unsupported claim / probably wrong. For every sentence in the third group, write one line on what the correct claim might be. Text: [SUBJECT]
For auditing text you already wrote.
05 — Tips
Notes that take the result up a level
- Assigning the model a labelling job works far better than telling it not to lie. You are not suppressing a behaviour, you are replacing it with a measurable output.
- Run the same question on two different models. Where both agree the answer is usually safe; where they diverge it is usually invented.
- If the model drops the tags mid-conversation, paste the tag rules again with the word "FORMAT LOCK" above them.
06 — FAQ
Frequently asked about AI Errors
Why not just say "don't hallucinate"?
Because the instruction is unenforceable. The model has no internal signal that distinguishes knowing from plausibly reconstructing. Giving it an output format that separates the two is something it can actually follow.
Does this work on Claude and Gemini too?
Yes. Claude tends to label conservatively, Gemini more liberally. If you get too many [NOT SURE] tags on things you know are correct, change the definition to "no source I can name" rather than "not sure".
Is Perplexity more reliable here?
Perplexity's citations are real links, so the failure mode changes: the link exists but often does not support the sentence. The [INFERENCE] and [NOT SURE] tags catch that gap, which a citation count never does.
Next
You may also find these useful
ChatGPT Refuses to Answer? A Prompt That Opens the Refusal
Separates the cause of a refusal (safety filter, over-caution, vague request) and builds the right frame for each. It gets your work done without wandering.
Intermediate ChatGPTClaudePrompt for When the Model Forgets Your Instructions Mid-Chat
The model writes your rules back in its own words, then checks every reply against that list. Still working 40 messages later.
Intermediate ChatGPTClaudePrompt That Turns a Generic AI Answer Into Real Analysis
Bans empty lines like "it is important" and demands a claim, a counter-claim and the missing side for each point. Real analysis in three rounds.
Intermediate ChatGPTClaudePrompt to Fix AI Code: Catch Hallucinated APIs
Forces an inventory of every API the code uses, flags invented symbols and wrong signatures, and gives a verification plan before you run anything.
AdvancedThis page was prepared by the CopyPromt editorial team, first published 25 October 2026. Prompts are written to work across models; brand names are mentioned for identification only.