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