claim-auditlisted

# Claim Audit ## Purpose Locate facts Claude asserted in past conversations that the user then corrected. Root-cause each confirmed incident, then output a prevention ruleset — search-first triggers and verification gates — so the same class of error stops recurring. Optionally stage those rules into persistent memory, but only behind an explicit confirmation gate. ## Triggers Run this skill when the user says `run claim audit`, `hallucination audit`, `what did you get wrong`, `claim audit`, or otherwise asks to find Claude's past incorrect claims and build rules to prevent them. ## Preconditions - Run outside any Project for a global audit. Chats inside a Project are invisible to conversation search from outside it (and vice versa). If invoked inside a Project, state up front that coverage is limited to that Project's chats. - Required tools: `conversation_search`, `recent_chats`, `read_conversation`. Optional: `memory_user_edits` (staging step only), Python + file access (deep mode only). - Default audit window: last 30 days. The user may widen or narrow it. ## Lexicon `conversation_search` is a literal keyword match, not semantic. Run exactly one search per term below — never merge terms into a single query, and keep phrasing as the content words users actually type when correcting. | Term (one search each) | |---| | actually it's | | that's wrong | | doesn't exist | | no such | | not a real | | hallucinat | | made that up | | outdated | | wrong version | | deprec