Self-Improvement Skill

Capture feedback and corrections to continuously enhance repository knowledge.

Philosophy

Knowledge gained during work should be captured immediately, not lost. Documentation updates go hand-in-hand with the work that revealed them — include them in the current branch, not separate PRs.

Execution Flow

Phase 1: Classify the Feedback

See references/classification.md for the full signal → classification → target mapping and decision tree for finding the right file.

Short summary: Corrections fix existing docs. New patterns go to module CLAUDE.md or skills. Anti-patterns go to root CLAUDE.md. Workflows with 5+ steps become skills.

Phase 2: Confirm Before Applying

Always use AskUserQuestion before making updates. Present:

1. Classified feedback type and proposed target location
2. Preview of the proposed change
3. Alternatives if uncertain

Phase 3: Apply Update

CLAUDE.md update: Read the file → find the appropriate section → add content maintaining existing style and structure.

Skill update: Read the existing SKILL.md → identify the relevant section → add or modify → update references if needed.

New skill: Create SKILL.md with proper frontmatter, add to the relevant agent's skills: list in .claude/agents/, add to .claude/settings.json allow list.

Correction: Find all occurrences of incorrect information → update each → verify no broken references result.

If uncertain about placement, ask rather than guess. If the target file doesn't exist, ask whether to create it or find an alternative location. If the update would be substantial (rewrites a section), preview the full change and consider whether it warrants a separate skill.

Integration with PR Workflow

Documentation updates are part of the current work. Group related code and doc changes together in the same commits. Proactively suggest documentation updates for new patterns discovered during implementation.

Proactive Invocation

Invoke this skill when:

• Completing implementation work where new patterns emerged
• The same concept has been explained twice (capture it)
• The user corrects an approach mid-task
• Undocumented behavior is discovered

Cross-References

• references/classification.md — Feedback classification table and target location decision tree