The Meta-Skill: Autonomous Skill Generator

Before writing any files, present a brief design to the user:

Skill Design: {skill-name}

Purpose: {what it does}
Trigger phrases: {when it activates}
Tools used: {shell, web_search, browser, etc.}
External deps: {API keys, CLI tools, or "none"}
Schedule: {cron expression or "on-demand"}

Generating skill...

Wait for confirmation only if the skill involves destructive actions or costs money. Otherwise, proceed immediately.

Step 3: Generate the SKILL.md

Read the format guide at {baseDir}/references/skill_format_guide.md for the complete SKILL.md specification. Read example skills at {baseDir}/references/example_skills.md for patterns and inspiration.

Generate a complete SKILL.md following these rules:

Frontmatter rules:

• name: kebab-case, descriptive, unique. Check existing skills first: ls ~/.openclaw/workspace/skills/ to avoid name collisions.
• description: THIS IS CRITICAL. Write it like you're describing the task to a coworker in chat. Include the actual nouns and verbs users will type. Include both what the skill does AND when to use it. This is the primary trigger mechanism — if the description doesn't match user language, the skill will never activate.
• Include metadata.openclaw.requires.env if API keys are needed.
• Include metadata.openclaw.requires.bins if CLI tools are needed.

Body rules:

• Write instructions in imperative form ("Fetch the data", not "You should fetch")
• Be specific: include exact API URLs, curl commands, file paths
• Include error handling: what to do if an API is down, rate limited, etc.
• Include output formatting: exactly how to present results to the user
• Keep it under 300 lines. If longer, split into reference files.
• Do NOT include "When to Use" sections in the body — that belongs in the description frontmatter only.
• Include example interactions showing input/output pairs.

Step 4: Generate Helper Scripts (if needed)

If the skill requires:

• Data processing that's complex or repetitive
• API calls with authentication headers
• File format parsing (JSON, CSV, XML)
• Any logic that would be unreliable as natural language instructions

Then create helper scripts in scripts/ directory. Scripts should be:

• Self-contained (no external dependencies beyond standard lib when possible)
• Executable (chmod +x)
• Well-commented
• Referenced from SKILL.md using {baseDir}/scripts/script_name.sh

Prefer Python for data processing, Bash for simple shell tasks.

Step 5: Write Files to Workspace

Create the skill directory and files:

# Create skill directory
mkdir -p ~/.openclaw/workspace/skills/{skill-name}

# Write SKILL.md
cat > ~/.openclaw/workspace/skills/{skill-name}/SKILL.md << 'SKILLEOF'
{generated content}
SKILLEOF

# If scripts are needed:
mkdir -p ~/.openclaw/workspace/skills/{skill-name}/scripts
cat > ~/.openclaw/workspace/skills/{skill-name}/scripts/{script}.sh << 'SCRIPTEOF'
{script content}
SCRIPTEOF
chmod +x ~/.openclaw/workspace/skills/{skill-name}/scripts/{script}.sh

Step 6: Validate

After writing files, run the validation script:

bash {baseDir}/scripts/validate_skill.sh ~/.openclaw/workspace/skills/{skill-name}/SKILL.md

Also manually verify:

# Check the file exists and has valid frontmatter
head -20 ~/.openclaw/workspace/skills/{skill-name}/SKILL.md

# Verify frontmatter has required fields
grep -A5 "^---" ~/.openclaw/workspace/skills/{skill-name}/SKILL.md | head -10

If validation fails, fix the issue and re-write the file.

Step 7.5: Voice Confirmation (if ElevenLabs is configured)

If the ELEVENLABS_API_KEY environment variable is set, generate a spoken confirmation:

bash {baseDir}/scripts/voice_confirm.sh "{skill-name}" "{example trigger phrase}"

This sends an audio message to the user confirming the skill was deployed. If ELEVENLABS_API_KEY is not set, skip this step silently.

Step 7: Test

Test the skill by checking it's properly installed:

# Check if skill appears in the skills list
openclaw skills list 2>&1 | grep {skill-name} || echo "Skill deployed. Will appear in next session or when skills watcher picks it up."

# List existing skills to confirm
bash {baseDir}/scripts/list_existing_skills.sh

Tell the user:

Skill "{skill-name}" deployed!

Location: ~/.openclaw/workspace/skills/{skill-name}/
Files: SKILL.md{, scripts/...}
Status: The skill will be available in your next session,
   or immediately if skills watcher is enabled.

Test it by saying: "{example trigger phrase}"

Want me to modify anything?

Step 8: Iterate

If the user wants changes:

• Read the existing SKILL.md
• Make targeted edits (don't regenerate from scratch unless major changes)
• Re-validate
• Confirm changes

Skill Editing Commands

• "edit skill {name}" — Modify an existing generated skill
• "delete skill {name}" — Remove a generated skill

  rm -rf ~/.openclaw/workspace/skills/{name}
  

• "list my skills" — Show all workspace skills

  bash {baseDir}/scripts/list_existing_skills.sh
  

• "show skill {name}" — Display the SKILL.md content

  cat ~/.openclaw/workspace/skills/{name}/SKILL.md
  

• "test skill {name}" — Trigger the skill with a test prompt

Quality Standards for Generated Skills

Every generated skill MUST:

1. Have a precise description that includes trigger words users actually type
2. Include error handling — what happens when things go wrong
3. Have example interactions — at least 2 input/output examples
4. Be self-contained — not depend on other custom skills
5. Use {baseDir} for any self-references to scripts or assets
6. Include security notes if handling sensitive data (API keys, passwords)
7. Be idempotent — running it twice shouldn't break anything

Available Tools Reference

When designing skills, these are the tools the generated skill can instruct the agent to use:

• shell/exec: Run any shell command. Most powerful tool. Use for: API calls (curl), file manipulation, running scripts, git, etc.
• web_search: Search the web. Use for: finding current info, research.
• browser: Full browser control (navigate, click, fill forms, screenshot). Use for: web scraping, form filling, site monitoring.
• filesystem (read/write): Read and write files on the local machine. Use for: config files, logs, data storage, note-taking.
• cron: Schedule recurring tasks via openclaw.json config.
• notifications: Send system notifications for alerts and reminders.

Read {baseDir}/references/available_tools.md for detailed tool documentation. Read {baseDir}/references/common_patterns.md for reusable skill patterns.

Anti-Patterns to Avoid in Generated Skills

• Do NOT use marketing copy in the body ("This amazing skill will revolutionize...")
• Do NOT put "When to Use" sections in the body (belongs in description frontmatter)
• Do NOT write vague instructions ("process the data appropriately")
• Do NOT skip error handling
• Do NOT hardcode API keys or secrets
• Do NOT create skills over 500 lines without reference file splits
• Do NOT assume tools the user hasn't enabled
• Do NOT use unicode bullets in frontmatter (use ASCII only in YAML)