Claim Audit — Security skill for Claude Code

claim-audit

A Claude skill that audits AI-written text claim by claim: quotes each checkable claim, tags it with one of ten hallucination patterns, verifies it against primary sources, and returns a verdict, a flagged table, and a fix-it prompt to send back to the model.

When a language model lacks reliable information it doesn't stop; it guesses, fluently and with confidence. This skill reads AI output the way a skeptical peer reviewer reads a manuscript, and hands back something you can act on in two minutes.

How it works

**0. Interrogate first, if the AI session is still open.** Send the model five chain-of-thought follow-ups at once — *cite your sources; which are most relevant; where are you guessing versus knowing; what should I verify; where could you be wrong* — and audit its answers too. The skill knows what those answers reliably do: "cite your sources" asked after the fact returns source-shaped categories rather than documents, and "where could you be wrong" returns generic hedges rather than the claim at risk. The self-report is a map of where to spend checks, not evidence.

**1. Inventory the claims.** Every statement that could be false — facts, numbers, names, dates, citations, quotations, rules, product features, causal claims — quoted exactly. Paraphrasing hides the tell: "23.7%" and "roughly a quarter" are different claims.

**2. Classify each claim against ten patterns.**

#	Pattern	The tell
1	Fabricated fact	A specific event, law or entity asserted with confidence and no source
2	Fabricated citation	Prestigious author or journal, no DOI, link, volume or page; or a real journal with a paper it never published
3	False precision	A decimal-point statistic with no methodology, population or year
4	Misattribution	A real idea, quote, rule or finding assigned to the wrong person or agency
5	Timeline error	"Recently", "the latest": a concrete event filling a gap about the current state