Catch Claude Making Things Up: The Free Lie Detector Prompt
One copy-paste prompt that splits any Claude answer into single claims, makes each one show where it came from, opens the real sources to check they actually say it, and hands you a clean answer with only what checks out. The full prompt, how to use it, the study behind it, and the catch
AI answers sound sure of themselves even when a detail is wrong. A made-up number, a study that doesn't exist, a quote nobody said: it all reads the same as the true parts. This free prompt turns Claude into its own lie detector. It checks the answer line by line, makes every claim show where it came from, flags anything it can't prove, and opens the real sources to make sure they're legit.
Here's the full prompt, how to use it, the study behind it, and the catch.
Why you need it
Researchers at Carnegie Mellon (Cash and Oppenheimer, published in Memory & Cognition in 2025) tested ChatGPT, Gemini and Claude Sonnet and Haiku against people. The chatbots overestimated how well they did, and unlike people, they didn't adjust after doing badly. On a sketch-guessing game, Gemini got fewer than 1 of 20 right on average, then guessed it had gotten about 14. Confidence isn't proof, so make the answer show its proof.
What it does
- Splits the answer into single claims, one fact per line.
- Gives each claim a source (title, publisher, link) or marks it UNVERIFIED. It is told never to invent a source.
- With web search on, opens every source and tags each claim LEGIT, MISMATCH or NOT FOUND.
- Shows a table with the problem claims at the top.
- Rewrites a clean answer with only the claims that checked out, plus a short list of what it couldn't verify.
The lie detector prompt
Run the lie detector on your last answer (or on the text I paste below). Follow these steps exactly. 1. SPLIT. Break the answer into separate claims, one fact per line. A claim is anything that could be true or false: a number, a date, a name, a quote, a rule, a "studies show", a cause and effect. Skip pure opinions and advice, but list any facts they lean on. 2. SOURCE. For each claim, give the source it came from: the page title, who published it, and the link. If you don't have a real source you can point to, write UNVERIFIED. Never make up a source, a link, a quote, a page number, a study, an author or a statistic. Saying "I don't know where this comes from" is always better than a source that might not exist. 3. CHECK. If you can search the web, open every source you listed and read it. Then tag each claim: - LEGIT: the source is real, it's trustworthy for this kind of claim, and it actually says this. - MISMATCH: the source is real but says something different (give what it really says). - NOT FOUND: the link is dead, the page doesn't exist, or you can't find the claim on it. If you can't open web pages, tag every sourced claim NOT CHECKED instead, and tell me to turn on web search. If you find a better source while checking, you can use it, and say that you swapped it. 4. FLAG. Show me a table: claim | source | tag. Put every UNVERIFIED, MISMATCH, NOT FOUND and NOT CHECKED claim at the top. 5. CLEAN ANSWER. Rewrite the answer using only the LEGIT claims. Fix the MISMATCH claims to match what the source really says, and only keep them if the fixed version still checks out. Leave out everything else. Don't add new facts in this step. 6. WHAT I COULDN'T VERIFY. End with a short list of the claims you removed or couldn't check, so I know what to look up myself. Rules: be strict. If you're not sure, it's UNVERIFIED. Don't soften a tag to make the answer look better, and don't apologize. Just give me the table, the clean answer and the list.
Free, for your email
Get the full setup steps + the copy-paste setup prompt
The rest of this guide (3 more sections) opens right here. The setup prompt, one paste into Claude that sets it all up, comes by email only. Use your real address.
50,000+ follow where these guides come from. No spam, unsubscribe in one click.