skill-factory（技能工厂）

Skill Creator

Build, refine, and publish OpenClaw skills. Supports six modes.

Modes at a Glance

Mode 1: Create

Build a skill from scratch in 6 steps.

Step 1 — Understand

Clarify before writing a single line:

• What does this skill do that no existing skill does?
• Who triggers it and when? (the description field drives triggering)
• What CLI tools, APIs, or files does it need?
• What's the output format?

Run scripts/analyze_patterns.py --query "<skill concept>" to see if relevant patterns already exist.

Step 2 — Plan

Write a one-paragraph spec covering: trigger conditions, happy path, error cases, output format. Confirm with user if uncertain.

Step 3 — Init

Scripts are bundled in scripts/ — no external path needed:

# From your workspace skills directory:
python3 $(openclaw skills info skill-creator --json 2>/dev/null | python3 -c "import json,sys; print(json.load(sys.stdin).get('path',''))")/scripts/init_skill.py \
  <skill-name> \
  --path ~/.openclaw/workspace/skills/ \
  --resources scripts,references \
  --examples

Or locate the skill dir and use relative path:

SKILL_DIR=$(dirname $(find ~/.openclaw/workspace/skills ~/.nvm -name "init_skill.py" 2>/dev/null | head -1))
python3 "$SKILL_DIR/init_skill.py" <skill-name> --path ~/.openclaw/workspace/skills/ --resources scripts,references

This creates:

<skill-name>/
  SKILL.md          # Edit this
  scripts/          # Helper scripts
  references/       # Reference docs, cheat sheets
  _meta.json        # Auto-populated on publish

Step 4 — Write SKILL.md

Frontmatter rules:

---
name: my-skill-name          # lowercase-hyphen, max 64 chars
description: "One sentence: what it does AND when to use it. Include trigger phrases."
---

Body structure:

# Skill Title

Brief one-liner.

## Quick Start
[Most common usage — 3-5 lines max]

## Commands / Recipes
[Concrete examples with real output]

## Reference
[Full option tables, edge cases, advanced usage]

Progressive disclosure rules:

• Frontmatter: always loaded (~100 words) — make it count
• Body: loaded on trigger (<500 lines) — stay under limit
• Bundled resources: loaded on demand — put verbosity here

Step 5 — Package

# package_skill.py is bundled in this skill's scripts/ directory:
SKILL_SCRIPTS="$(dirname "$(find ~/.openclaw/workspace/skills/skill-creator ~/.nvm -name "package_skill.py" 2>/dev/null | head -1)")"
python3 "$SKILL_SCRIPTS/package_skill.py" ~/.openclaw/workspace/skills/<skill-name>

Validates structure, outputs <skill-name>.skill zip.

Step 6 — Iterate

Run evals (Mode 2) → identify failures → update SKILL.md → re-package → repeat.

Mode 2: Eval

Measure skill quality against defined expectations.

Setup

Create evals/evals.json:

[
  {
    "id": "basic-create",
    "prompt": "Create a skill that sends a Slack message",
    "expected_output": "SKILL.md with slack-notifier name and working command",
    "assertions": [
      "contains SKILL.md frontmatter with name and description",
      "contains at least one bash command example",
      "description includes trigger phrases"
    ]
  }
]

Eval Run

For each eval case:

1. Execute the prompt using current skill
2. Grade against assertions (pass/fail per assertion)
3. Log result to evals/runs/<timestamp>.json

Run Report Format

{
  "skill": "skill-creator",
  "version": "1.0.0",
  "timestamp": "2026-02-22T03:00:00Z",
  "pass_rate": 0.85,
  "cases": [
    { "id": "basic-create", "passed": true, "assertions_passed": 3, "assertions_total": 3 }
  ]
}

Mode 3: Improve

Iterate on an existing skill using eval feedback.

Improvement Loop

1. Run evals → identify failing assertions
2. Read current SKILL.md
3. Draft changes targeting failures
4. Write new version (increment semver in _meta.json)
5. Re-run evals → confirm pass rate improved
6. Update history.json

history.json

Track all versions at evals/history.json:

[
  {
    "version": "1.0.0",
    "parent": null,
    "expectation_pass_rate": 0.70,
    "is_current_best": false,
    "notes": "Initial version"
  },
  {
    "version": "1.1.0",
    "parent": "1.0.0",
    "expectation_pass_rate": 0.85,
    "is_current_best": true,
    "notes": "Improved trigger description, added Synthesize mode"
  }
]

Mode 4: Benchmark

Blind A/B comparison of two skill versions.

Process

1. Run identical eval suite against version A and version B
2. Collect raw outputs without labels
3. Compare blind (no version labels) → pick winner per case
4. Reveal versions, compute delta
5. Recommend: keep A, adopt B, or cherry-pick specific cases

Benchmark Output

Version A: 1.0.0  pass_rate=0.70
Version B: 1.1.0  pass_rate=0.85
Delta: +0.15 (B wins)
Regressions: 0
Recommendation: Adopt B

Mode 5: Analyze Patterns

Scan installed skills to extract reusable building blocks.

python3 ~/.openclaw/workspace/skills/skill-creator/scripts/analyze_patterns.py \
  --scan-dirs ~/.openclaw/workspace/skills/,~/.nvm/versions/node/v22.22.0/lib/node_modules/openclaw/skills/ \
  --output ~/.openclaw/workspace/skills/skill-creator/references/patterns.md

What it extracts:

• Trigger phrases — common description keywords that activate skills
• Tool patterns — CLI tools, APIs, Docker patterns used across skills
• Output formats — JSON schemas, markdown templates, log formats
• Structural patterns — how skills organize commands/recipes
• Error handling patterns — retry logic, circuit breakers, fallbacks

See references/patterns.md for the current extracted pattern library.

Mode 6: Synthesize from Patterns

Build a new skill scaffold by combining patterns from the library.

Usage

When asked to create a skill in a domain that resembles existing skills:

1. Run Analyze Patterns first
2. Query references/patterns.md for relevant patterns
3. Compose a SKILL.md that combines:

- Best trigger phrases from similar skills - Relevant tool/API patterns - Appropriate output format - Error handling from most robust similar skill

Example

"Create a skill for Twitter scraping":

• Pull trigger phrases from reddit-scraper
• Pull CDP/browser patterns from fast-browser-use
• Pull output format (JSON array) from crypto-market-data
• Synthesize into twitter-scraper/SKILL.md

Skill Anatomy Quick Reference

<skill-name>/
  SKILL.md           # Required: frontmatter + body
  scripts/           # Helper Python/bash scripts
  references/        # Cheat sheets, API docs, schemas
  assets/            # Images, templates
  evals/
    evals.json       # Test cases
    runs/            # Eval run results
    history.json     # Version history
  _meta.json         # Publishing metadata

_meta.json template:

{
  "ownerId": "",
  "slug": "skill-name",
  "version": "1.0.0",
  "publishedAt": null
}

Publishing to OpenClaw Community

Registry: clawhub.com — use the clawhub CLI (already installed).

# 1. Login (opens browser once)
clawhub login

# 2. Publish directly from skill folder — no .skill zip needed
clawhub publish ~/.openclaw/workspace/skills/<skill-name> \
  --version 1.0.0 \
  --changelog "Initial release"

# 3. Or sync all workspace skills at once:
clawhub sync --workdir ~/.openclaw/workspace --dir skills

1. Ensure _meta.json has correct slug and version
2. Run full eval suite — pass rate must be ≥ 0.80
3. clawhub login (one-time browser auth)
4. clawhub publish <skill-folder>
5. Verify at clawhub.com/skills/<slug>

Quality bar for publishing:

• [ ] Description triggers correctly (test with 3+ natural phrasings)
• [ ] At least 3 concrete command examples with real output
• [ ] Error cases documented
• [ ] Eval pass rate ≥ 0.80
• [ ] _meta.json complete