Skill Improver

Retrospect and improve Skills/Workflows based on real execution artifacts, making the skill chain more stable with use.

Input (pass path only)

• run_dir: runs/<workflow>/active/<run_id>/ (or openspec/changes/<change-id>/)

Output

• improvements.md: Improvement suggestions (can be directly converted to change tasks)
• Optional: Minimal patch to relevant SKILL.md (only change what's necessary)

Process

1. Read logs/state.json and logs/events.jsonl, extract:
   • Failure points (error types, frequency, missing context)
   • Time bottlenecks (repetitive steps, ineffective searches, excessive context)
   • Manual confirmation points (whether too early/too late/missing)
2. Check artifact contracts:
   • Whether each step has persisted output
   • Whether only paths are passed instead of content
   • Whether there are inconsistent naming/paths
3. Output improvement suggestions (sorted by benefit):
   • Prompt/trigger word optimization
   • I/O contract completion (add required fields, fix artifact file names)
   • Error handling and rejection strategies
   • Suggest new deterministic scripts (only when necessary)

Output Template (recommended)

• Overview: Core issues from this run
• Top Issues (high → low): Each includes "symptom/root cause/fix suggestion/related skill/suggested change point"
• Suggested Minimal Change List: Listed by file