Skill Workshop: A Review Step Before Agent Skills Go Live

OpenClaw has shipped Skill Workshop, a structured feature for creating and refining reusable agent skills with a human review step built in before anything goes live.

Skills in OpenClaw are how agents learn reusable procedures — invoice follow-up checklists, release routines with validation steps, repo health workflows with scripts and templates. But unlike a normal file edit, a skill doesn’t just answer one question: it changes how future work is done. A bad skill can quietly corrupt every subsequent run.

Skill Workshop addresses this by inserting a proposal layer between the agent’s draft and live skill execution:

Proposal First — When an agent creates or revises a skill through Skill Workshop, it starts as PROPOSAL.md, not SKILL.md. The proposal carries the draft instruction, support files, review state, and everything needed to apply or reject it cleanly. It doesn’t run yet.

Shape It Before You Keep It — The Control UI surfaces proposed skills as reviewable work in Board view (full workshop with pending/applied/rejected/stale states) and Today view (one-at-a-time fast pass with a concrete approve/edit/skip decision).

Tweak Is Where It Matters — Generated skills are often almost right. Skill Workshop turns “close but not quite” into a revision conversation: adjust the wording, add a safer fallback, convert a script to a template. The proposal is the object being edited, history intact.

Skills can also carry support files — response templates, example logs, deployment scripts — which are reviewed alongside the main instruction file.

Skill Workshop is available now via the OpenClaw Control UI and CLI (openclaw skills).