Prompt libraryChecking AI answers
Prompts that expose weak claims, test for flattery and prepare a verification list.
These prompts help you decide what to verify. They do not verify anything themselves: the same assistant can be confidently wrong about its own weak spots.
Combine them with an independent second model and, for anything that matters, a source you open yourself.
Here is an answer you gave earlier:
[paste]
List the three claims most likely to be wrong or out of date, and for each say what source would settle it.Why it works. Focuses your checking time.
Then check. Check those three first.
Before answering, tell me whether my question contains any assumption that might be false: "[question]". Then answer it.Why it works. Catches false premises models would otherwise accept.
Then check. Verify the premise independently.
Answer the opposite question: "[reverse of your question]". Then tell me whether your two answers are consistent.Why it works. Inconsistency across framings suggests flattery.
Then check. Compare for contradictions.
Rewrite your answer in three labeled groups: things you are confident are well established, things you inferred, and things you are guessing. If you are unsure which a statement is, put it under guessing.Why it works. Forces a distinction models usually blur.
Then check. Treat the labels as a hint, not proof.
For each factual claim in your answer, name a type of source that would contain it. Do not invent titles, authors, links or numbers; if you can't point to a kind of source, say so.Why it works. Discourages fabricated references.
Then check. Open real sources yourself.
Another assistant answered this question: "[question]" with:
[paste]
Where do you agree, where do you disagree, and what does the answer assume that it didn't state?Why it works. A different model's critique is a cheap cross-check.
Then check. Resolve differences with a source.