retro— agent skill

Run a retrospective on the work just completed.

When to Use

• After completing a non-trivial task (new tool, bug fix, refactor)
• After a task that required rework or user corrections
• After the self-correction protocol fired
• When the user explicitly asks

Workflow

1. Review What Happened

• Check git diff for what changed
• Scan conversation for user corrections, stops, or direction changes
• Count: how many times was the approach adjusted?

2. Categorize

3. Write Lessons (for Corrected/Reworked only)

Write to auto-memory (feedback type) with:

• name: Short descriptive title
• description: One-line summary for future relevance matching
• Content:
• Trigger: What happened
• Wrong: What I did and why it was wrong
• Right: What should have been done instead
• Rule: The general principle (one sentence, reusable)
• Area: Which part of the project this applies to (tools, endpoints, docs, publishing, or "all")

4. Check for Promotions

Review existing auto-memory feedback entries. If the same lesson pattern appears 3+ times:

1. Propose adding as a permanent rule to the Known Patterns & Pitfalls table in CLAUDE.md
2. Get user approval before modifying CLAUDE.md
3. Once promoted, note the promotion in the auto-memory entry

5. Check if Skills Need Updates

If a lesson reveals a gap in a skill workflow (e.g., /add-tool missed a step, /verify should check something extra):

1. Identify which skill file needs updating
2. Propose the specific change
3. Get user approval before modifying

Output Format

## Retrospective

**Task:** [what was done]
**Outcome:** Clean / Corrected / Reworked
**Adjustments:** [N times the approach was changed]

### Lessons Learned
[If Corrected/Reworked — list each lesson with the rule]

### Promotions
[If any feedback pattern hit 3+ occurrences — propose CLAUDE.md update]

### Skill Updates
[If any skill workflow gap was identified — propose the fix]

6. Update Issue Tracker

If bugs were found or fixed during this task:

• Update ISSUES.md with any new issues discovered (assign ISS-NNN IDs)
• Update status of fixed issues to FIXED with resolution details

Rules

• Be honest about mistakes — the point is improvement, not ego preservation
• Every lesson must have a concrete, actionable Rule
• Don't write lessons for things that went smoothly — noise drowns signal
• Focus on lessons that are reusable across future tasks, not one-off fixes
• If the same mistake happened because a CLAUDE.md rule wasn't followed, note that — the rule exists, the issue is compliance
• Bugs found during retro MUST be logged to ISSUES.md — don't rely on memory