AI agent skill · by @clawdotnet

Meta Skill Creator

Use this meta-skill instead of answering directly only when the current user explicitly asks to create, compose, synthesize, or propose a new meta-skill that orchestrates multiple existing skills. It uses multi-skill orchestration for intent clarification, optional history mining, trigger-collision checks, linting, smoke/runtime gates, preview, and optional proposal persistence. Do not use it for creating a normal standalone skill, asking how meta-skills work, analyzing pasted skill lists, or discussing existing meta-skills.

Archived 2026-10-23 · Source updated 2026-10-10 · License: MIT. Check the original repository for the latest version.

SKILL.md instructions

# Meta-Skill Creator

Safeguarded DAG that synthesizes a new bundled meta-skill from observed skill
co-occurrence patterns + user description of the desired workflow. It now
separates preview-only, persisted-proposal, and fully gated modes so lightweight
requests do not pay for persistence or smoke testing. The workflow separates
generic skill creation from meta-skill composition, checks trigger collisions,
classifies operational risk, and previews the proposal before optional
persistence.

Output is a SKILL.md candidate written to `~/.openclaw/proposals/<id>/`.
By default it is not auto-loaded; run `openclaw skills meta-runs proposals accept <session-id> --proposal <id>`
to enable. If the operator has enabled the auto-propose `auto_enable` setting,
this manual path also runs the same conservative static safety preflight used by
cron/dream auto-propose and may promote a low-risk gated proposal immediately.

## Fallback

If creator's pipeline fails at any step, **report the failure verbatim** to the
user:

1. State which step failed (e.g. "harvest", "lint")
2. Quote the error message from the orchestrator's structured log
3. Stop. Do NOT improvise.

Do NOT:
- Claim a proposal was written unless you have verified it by reading
  `~/.openclaw/proposals/<id>/SKILL.md` with the `read_file` tool
- Invent file paths, proposal IDs, or skill names that you have not seen
  in the orchestrator's actual output
- "Manually run" the individual skills as a recovery — that bypasses
  the validation gates the user explicitly opted into

If the user wants to retry, suggest they re-issue the request after the
underlying error is resolved (often a sandbox or provider issue), not a
manual workaround.