auto-skill-extractor（自动技能提取器）

Auto-Skill Extractor

Turn your agent's work into reusable skills. Automatically.

When to Use

✅ You run complex subagent tasks repeatedly ✅ You want to build a skill library without manual authoring ✅ You run multi-domain tasks (files + system + web) ✅ You want your agent to learn from its own patterns

When NOT to Use

❌ Simple 1-2 tool tasks (not worth skilling) ❌ One-off exploratory work ❌ You prefer manually authoring every skill

Quick Start

1. Install

clawhub install auto-skill-extractor

2. Create Directories

mkdir -p skills/auto-draft skills/auto skills/manual

3. Wire Into Your Agent

Add to AGENTS.md after subagent completion:

# Auto-skill extraction trigger
import subprocess
import json

# Write trigger input
trigger_data = {
    "completion_status": "success",
    "tool_calls": tool_call_count,  # from subagent result
    "transcript_summary": brief_summary,  # Keep brief, avoid secrets
    "session_id": session_key,
    "multi_domain": True  # if applicable
}

# RECOMMENDED: Pipe via stdin (no file on disk)
result = subprocess.run(
    ["python3", "scripts/auto-skill-trigger.py"],
    input=json.dumps(trigger_data),
    capture_output=True,
    text=True
)

# ALTERNATIVE: File-based (delete immediately after)
# with open("/tmp/trigger.json", "w") as f:
#     json.dump(trigger_data, f)
# result = subprocess.run(
#     ["python3", "scripts/auto-skill-trigger.py", "/tmp/trigger.json"],
#     capture_output=True, text=True
# )
# os.remove("/tmp/trigger.json")  # SECURITY: Delete after use

output = json.loads(result.stdout)
if output.get("action") == "extract":
    print(f"🔄 Created DRAFT skill: {output['skill_name']}")

4. Use It

Run a subagent with complex work:

Spawn subagent to analyze codebase...
- Read 5 config files ✓
- Check processes ✓
- Write summary report ✓

Result: skills/auto-draft/codebase-analyzer-abc123/ created automatically

How It Works

Step 1: Trigger Evaluation

After every subagent completion:

Step 2: Complexity Scoring

Base:    tool_calls × 0.7  (max 5 pts)
          3 tools = 2 pts, 5 tools = 4 pts

Bonus:   +2  multi-domain (files + system + web)
         +2  error recovery (retry logic)
         +1  fail-then-succeed

Threshold: 4 points = extract

Step 3: DRAFT Creation

skills/auto-draft/my-skill-abc123/
├── SKILL.md      ← Template with metadata
└── meta.json     ← Invocation tracking

Step 4: Evaluation Period

• Use the DRAFT skill 3 times successfully
• Each use logged in meta.json
• After 3rd use → auto-promoted to skills/auto/

Step 5: ACTIVE Status

Promoted skills are:

• Visible in /skills auto list
• Ready for manual completion
• Versioned and tracked

Configuration

Edit scripts/auto-skill-trigger.py:

COMPLEXITY_THRESHOLD = 4    # Lower = more drafts, more curation
MAX_QUEUE_SIZE = 50         # Pending extraction limit
PROMOTE_THRESHOLD = 3       # Invocations before promotion

Manual Control

Force Extraction

Ignore thresholds:

#skill: force

List Drafts

python3 scripts/skill-lifecycle.py drafts

Promote Early

python3 scripts/skill-lifecycle.py promote my-skill-name

Archive Stale Drafts

python3 scripts/skill-lifecycle.py process
# Removes drafts unused for 7+ days

Safety

• ✅ Collision detection — Won't overwrite existing skills
• ✅ Path sanitization — ../../../etc → blocked
• ✅ Atomic promotion — Copy → verify → move → delete
• ✅ Queue limits — Max 50 pending extractions
• ✅ Return value checks — Errors logged, not silent

Verification

Check extraction worked:

# See recent DRAFTs
ls -la skills/auto-draft/

# Check extraction queue
cat scripts/skill-extraction-queue.json

# View specific skill
cat skills/auto-draft/my-skill-abc123/SKILL.md

Pitfalls

Architecture

Subagent completes
    ↓
auto-skill-trigger.py
    ↓
Score complexity (0-10)
    ↓
If ≥ 4: Create DRAFT
    ↓
skill-lifecycle.py
    ↓
After 3 uses: PROMOTE → skills/auto/
    ↓
After 7 days: ARCHIVE

Related

• Hermes Agent — Original inspiration (Nous Research)
• Auto-skill pipeline case study — https://github.com/wahajahmed010/openclaw-hermes-adaptation
• Skill lifecycle management — Built into this package

References

• Hermes Adaptation: https://github.com/wahajahmed010/openclaw-hermes-adaptation
• Original Hermes: https://github.com/NousResearch/Hermes-Agent