reflect-and-learn

To save user corrections to CLAUDE.md, the skill captures behavioral feedback during a session, validates it semantically, and routes approved learnings to the most specific persistence target, such as a skill-level or project-level file. This ensures corrections permanently persist across sessions.

Semantic validation for learning capture is a mechanism that applies confidence scoring to user corrections to ensure only genuine behavioral improvements are kept. It filters out transient or invalid feedback before routing approved learnings to the appropriate CLAUDE.md file.

Yes, reflect-and-learn requires file system access to read and write CLAUDE.md files during project onboarding or periodic reviews. This access is necessary to route validated corrections to the correct skill-level, project-level, or global persistence target.

To manage rule conflicts when writing corrections to multiple CLAUDE.md files, the skill detects and surfaces conflicting rules before writing. This conflict management mechanism preserves consistency across global, project-level, and skill-level persistence targets.

Yes, you can discover recurring workflow patterns by analyzing historic session corrections. The skill optionally analyzes past sessions to suggest new skills or identify repeating behavioral patterns, extending learning capture beyond immediate session feedback.