auto-skill-extractor#

Automatically learn from your AI's work and turn repeated subagent tasks into reusable skills.

Description#

Watches subagent completions, detects patterns worth reusing, and creates draft skills automatically. After 3 successful re-invocations, drafts are promoted to active skills.

How It Works#

1. Trigger: After every subagent completion, the main agent calls scripts/auto-skill-trigger.py
2. Score: Complexity is calculated (tool calls × 0.7, +2 for multi-domain, +2 for error recovery)
3. Extract: If score ≥ 4 and ≥ 3 tool calls → create DRAFT skill in skills/auto-draft/
4. Evaluate: After 3 successful re-invocations → promote to skills/auto/
5. Archive: Drafts unused for 7+ days → archived

Integration#

Add to AGENTS.md or your agent's completion handler:

python

# After subagent completion
import subprocess, json

trigger_input = {
    "completion_status": "success",  # or "failed", "timeout"
    "tool_calls": tool_call_count,
    "session_id": session_key,
    "multi_domain": True,  # if task crossed domains (files + web + system)
    "transcript_summary": brief_summary  # keep brief, no secrets
}

# Method 1: Stdin (recommended, no file on disk)
result = subprocess.run(
    ["python3", "scripts/auto-skill-trigger.py"],
    input=json.dumps(trigger_input),
    capture_output=True, text=True
)

# Method 2: File (delete immediately after)
with open("/tmp/trigger.json", "w") as f:
    json.dump(trigger_input, f)
result = subprocess.run(["python3", "scripts/auto-skill-trigger.py", "/tmp/trigger.json"], capture_output=True)
os.remove("/tmp/trigger.json")

Directives#

• #skill: <name> — Manually trigger extraction from last subagent
• #skill: force — Force extraction (ignore thresholds)

Commands#

bash

# List active auto-skills
python3 scripts/skill-lifecycle.py list

# List drafts under evaluation
python3 scripts/skill-lifecycle.py drafts

# Process promotions and archives
python3 scripts/skill-lifecycle.py process

# Manually promote a draft
python3 scripts/skill-lifecycle.py promote my-skill-name

Complexity Scoring#

FactorPointsExampleTool calls × 0.72-5 pts3 tools = 2, 5 tools = 4Multi-domain+2 ptsFiles + web + systemError recovery+2 ptsRetry on failure

Threshold: 4 points = creates DRAFT

Directory Structure#

text

skills/
├── auto-draft/          # Draft skills (under evaluation)
│   └── {name}-{hash}/
│       ├── SKILL.md
│       └── meta.json
├── auto/                # Promoted skills (ready to use)
│   └── {skill-name}/
│       ├── SKILL.md
│       └── promotion.json
└── manual/              # Hand-crafted skills

Configuration#

• COMPLEXITY_THRESHOLD = 4 — Lower = more drafts, more curation needed
• PROMOTE_THRESHOLD = 3 — Re-invocations before promotion
• ARCHIVE_AGE_DAYS = 7 — Days before unused drafts archived
• MAX_QUEUE_SIZE = 50 — Pending extraction limit

Safety#

• ✅ No sensitive data persisted (only tool counts, scores, timestamps)
• ✅ Session IDs are hashed, not stored raw
• ✅ Path traversal prevention on skill names
• ✅ Atomic promotions (copy → verify → move)
• ✅ Queue limits (max 50 pending)
• ✅ No transcript content written to disk

Companion Skills#

• subagent-orchestration — The subagent pattern that generates the work this skill extracts from
• skill-creator — For manually creating or improving skills