reflectlisted

# Reflect: Self-Improving Skills Scan the current conversation for signals (corrections, approvals, patterns) and propose updates to the skill files that were invoked during the session. Every correction should only need to happen once. ## Step 1: Identify Skills Used Scan the conversation for skill invocations. Look for: - Explicit `/skill-name` calls - Skill tool invocations (the Skill tool being called) - Agent delegations that loaded a skill If no skills were used this session, check whether the corrections apply to: - General behavior → suggest a CLAUDE.md or memory update instead - A specific skill that *should* have been used → note the routing gap - Neither → report "no learnings to extract" and stop ## Step 2: Extract Signals Scan the full conversation for three signal types: **HIGH confidence (explicit corrections):** - User said "no," "don't," "never," "stop," "wrong," "not that" - User rejected output and asked for a different approach - User provided a specific rule ("always do X," "use Y not Z") - User flagged an error, bad pattern, or bad output **MEDIUM confidence (validated patterns):** - User approved output without pushback ("yes," "perfect," "looks good," accepted without changes) - A non-obvious approach worked and the user confirmed it - User's positive reaction to a specific technique or format **LOW confidence (observations):** - Patterns that seemed to work but weren't explicitly validated - Ambiguous signals worth reviewing later **Not a si