rem-feedback

Capture typed feedback about skill mistakes so skills can learn and improve over time. Writes structured entries to ~/.claude/skill-feedback.jsonl tagged with one of six mistake types — misrouted (wrong skill fired / should have fired), wrong-output (false claim, bad advice), missed (skill didn''t flag an issue it should have), over-flagged (skill flagged a non-issue), context-ignored (skill ignored CLAUDE.md / memory / project conventions), stale (advice references deleted files / old APIs / outdated facts). Links each feedback entry back to the originating invocation in ~/.claude/skill-usage.jsonl by session_id + timestamp. Auto-promotes to /rem-skill analyze [name] when 3+ same-type flags accumulate for one skill. Distinct from rem-learn (which captures project-level lessons, not skill-system lessons) and rem-handoff (session state, not feedback). Use for "flag a skill mistake", "this skill got it wrong", "the skill should have X", "rem-copy misrouted", "rem-review missed the bug", "skill feedback", "rate the skill", "this wasn''t helpful", after any skill-assisted session where output disappointed.