skill-curator

name	skill-curator
description	Autonomous skill lifecycle manager. Reviews skill-compounder drafts, promotes high-confidence patterns, merges duplicates, archives low-quality drafts. Zero manual review required - runs on every session start.

Skill Curator

Autonomous skill lifecycle system. Works in partnership with skill-compounder to close the loop: observe -> draft -> review -> promote.

The Pipeline

1. You work normally (no manual action needed)
2. skill-compounder (Stop hook) detects successful patterns
3. Draft written to ~/.claude/skills-drafts/
4. skill-curator (SessionStart hook) reviews all drafts
5. Drafts are either:
   - PROMOTED to ~/.claude/skills/ (confidence >= 80, no duplicates)
   - MERGED with existing similar skill
   - ARCHIVED to ~/.claude/skills-drafts/archive/ (confidence < 50)
   - KEPT in drafts for next review (50-79, needs more evidence)

Promotion Criteria

A draft is auto-promoted when ALL of these are true:

Criterion	Threshold
Confidence	>= 80
Occurrences	>= 3 (or category is error-recovery)
Uniqueness	Not duplicate of existing skill
Age	Less than 14 days old
Category	One of: repeated-workflow, error-recovery, domain-pattern

Archive Criteria

A draft is auto-archived when ANY of these are true:

• Confidence < 50 (not reproducible)
• Age > 14 days without promotion (stale)
• Duplicate of existing promoted skill

Duplicate Detection

Before promoting, check against existing skills:

1. Name match - exact directory name
2. Category + description overlap - same category, 4+ keyword overlap

If duplicate found, either merge evidence into existing skill or archive.

Draft File Format

Drafts live in ~/.claude/skills-drafts/<name>.md:

---
name: pattern-abc123
description: "Procedural pattern: Read -> Edit -> Test -> Commit"
category: repeated-workflow
confidence: 85
occurrences: 7
source_session: s-xyz789
generated_at: 2026-04-12T03:45:00Z
status: draft
---

When promoted, status: draft becomes status: active.

Integration

• skill-compounder (Stop hook): Generates drafts from session observations
• self-learner: Provides error-pattern data that feeds skill-compounder
• canavar: Cross-trains all agents when skill is promoted
• layered-recall: Makes new skills queryable via Depth 1-3 pattern

Metrics

Tracked in ~/.claude/skill-evolution.jsonl:

• draft_generated_at
• promoted_at
• time_to_promotion
• usage_after_promotion
• retention_rate (still active after 30 days)

Safety

• No manual override needed - runs silently on session start
• No destructive actions - drafts/skills are moved, not deleted
• Rollback via git - all promoted skills are in git history
• High threshold - 80+ confidence prevents low-quality promotion