Evolve Skill

Evolve Skill

Analyze feedback and improve skills through structured evolution. This meta-skill enables skills to learn and improve over time.

Arguments

• $0 - Skill name to evolve (e.g., add-chart)
• Or special commands:
  • --all - Analyze all skills
  • --analyze-only {skill} - Just show analysis, no changes
  • rollback {skill} - Rollback to previous version
  • rollback {skill} {version} - Rollback to specific version
  • versions {skill} - List available versions

Mode 1: Evolve a Skill

Step 1: Gather Feedback

Read all feedback files for the target skill:

.claude/skills/.feedback/${$0}/*.yaml

Parse each feedback file and aggregate:

• Outcome distribution (success/partial/failed)
• Common strengths (patterns that appear multiple times)
• Common corrections (fixes that were needed repeatedly)
• Common suggestions (frequently requested improvements)

Step 2: Analyze Patterns

Categorize improvements by type:

1. Instructions - Steps that need clarification or reordering
2. Templates - Code patterns that need updating
3. Verification - Missing checklist items
4. Tools - Tool permissions that should change

Look for patterns:

• If 2+ feedback entries mention the same correction → likely needed
• If 2+ feedback entries suggest the same improvement → worth considering
• If success rate < 70% → skill needs significant improvement

Step 3: Propose Improvements

Present proposals to the user using AskUserQuestion:

Based on {N} feedback entries for '{skill-name}':
- Success rate: {X}%
- Common corrections: {list}
- Common suggestions: {list}

Proposed improvements:
1. {Improvement 1} - rationale from feedback
2. {Improvement 2} - rationale from feedback
...

Which improvements should we apply?

Options should include:

• Apply all proposed improvements
• Select specific improvements
• Skip (just update metadata)

Step 4: Apply Changes (with approval)

For each approved improvement:

1. Save current version to .versions/

   cp SKILL.md .versions/{current-version}.md
   

2. Edit SKILL.md with the improvement

   • Be precise with edits
   • Maintain skill structure and formatting
   • Add new sections if needed

3. Update metadata:

   version: {new-version}  # Bump appropriately
   last-evolved: {today's date}
   evolution-count: {previous + 1}
   

4. Update CHANGELOG.md:

   ## [{new-version}] - {date}
   
   ### Changed
   - {Description of change 1}
   - {Description of change 2}
   
   ### Feedback-driven
   Based on {N} feedback entries. Key patterns addressed:
   - {Pattern 1}
   - {Pattern 2}
   

Step 5: Commit Changes

Create a git commit with:

git add .claude/skills/${$0}/
git commit -m "evolve(${$0}): {brief description}

Based on {N} feedback entries:
- {Key improvement 1}
- {Key improvement 2}

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>"

Step 6: Report

Tell the user:

• Changes applied
• New version number
• Suggest testing the skill with /skill-name to verify

Mode 2: Rollback

rollback {skill} - Previous Version

1. Read current version from SKILL.md frontmatter
2. Find previous version in .versions/
3. Copy previous version to SKILL.md
4. Update metadata (decrement evolution-count)
5. Add CHANGELOG entry noting rollback
6. Commit with message: revert(${skill}): rollback to {version}

rollback {skill} {version} - Specific Version

1. Verify version exists in .versions/{version}.md
2. Copy that version to SKILL.md
3. Update metadata appropriately
4. Add CHANGELOG entry
5. Commit

Mode 3: List Versions

versions {skill}

1. List all files in .versions/ directory
2. Show version, date (from filename or metadata), and brief description
3. Indicate current version

Output format:

Versions for '{skill}':
  1.0.0 - 2026-02-02 - Initial release
  1.1.0 - 2026-02-15 - Added ErrorBoundary pattern (current)

Mode 4: Analyze All

--all

1. Scan all skill directories
2. For each skill with feedback:
   • Calculate success rate
   • Count pending feedback entries
   • Identify skills most in need of evolution

Output format:

Skill Evolution Status:
  add-chart      - 5 feedback entries, 60% success, needs evolution
  add-api-route  - 2 feedback entries, 100% success, healthy
  add-dashboard  - 0 feedback entries, no data
  ...

Recommend which skills to evolve first.

Version Numbering

Use semantic versioning:

• Patch (1.0.x) - Small fixes, clarifications

  • Typo fixes
  • Minor wording improvements
  • Adding missing details

• Minor (1.x.0) - New content, significant improvements

  • New sections added
  • Template code updated
  • Verification steps added

• Major (x.0.0) - Breaking changes (rare for skills)

  • Complete restructure
  • Argument changes
  • Fundamental approach change

Feedback Analysis Patterns

Detecting Instruction Issues

# If corrections mention "order" or "sequence"
corrections:
  - "Had to reorder the steps"
  - "Step 3 should come before Step 2"
→ Instruction ordering issue

Detecting Template Issues

# If corrections mention specific code
corrections:
  - "Fixed import order"
  - "Added missing ErrorBoundary"
→ Template needs updating

Detecting Missing Guidance

# If suggestions mention "add" or "include"
suggestions:
  - "Add section about error handling"
  - "Include z-index best practices"
→ Missing documentation

Verification Checklist

After evolution:

1. Previous version saved to .versions/
2. SKILL.md updated with improvements
3. Version bumped appropriately
4. last-evolved date updated
5. evolution-count incremented
6. CHANGELOG.md entry added
7. Git commit created
8. User notified of changes

After rollback:

1. Correct version restored
2. Metadata updated
3. CHANGELOG notes rollback
4. Git commit created

Converted and distributed by TomeVault — claim your Tome and manage your conversions.