Skill Workshop

Page scope: Skill Workshop is for creating and reviewing skills inside OpenClaw—not installing community skills from ClawHub. For installs see Install ClawHub skills safely and Create a skill tutorial.

Skill Workshop shipped in stable v2026.6.1. It gives you a governed workflow for skill creation: proposals land in a reviewable list, you can preview files, hand off revisions, and apply or reject changes through the skill_workshop agent tool—instead of letting the agent write skill files directly into production paths without oversight.

• Proposal lists — Skills start as proposals with clear status instead of silent overwrites.
• Revision handoff — Send work back for edits before anything goes live.
• Searchable file previews — Inspect skill contents before apply.
• Review states — Track pending, approved, or quarantined work.
• skill_workshop tool — Agent actions: apply, reject, or quarantine proposals under your policy.

• You want the agent to draft skills but a human (or stricter policy) approves before install.
• You run multiple agents or operators and need auditability for skill changes.
• You are moving from ad-hoc SKILL.md edits to a repeatable internal process.

For one-off personal skills you write yourself, the classic create-skill tutorial may be enough. Workshop adds governance when agents propose new capabilities.

1. Upgrade to v2026.6.1+: openclaw update and openclaw doctor --fix.
2. Confirm tools.profile and tools.allow expose skill-related tools your workflow needs (Configuration).
3. In Control UI or chat, ask the agent to propose a skill for a concrete task (e.g. a weekly report formatter).
4. Open the Workshop view (or proposal list) and preview generated files.
5. Apply if satisfied, reject with notes, or quarantine suspicious proposals.
6. Run one end-to-end test: invoke the new skill from a channel you trust.

Skill Workshop governs skill files. Workboard (same v2026.6.1 train) orchestrates multi-agent task runs on a board—useful when proposals spawn sub-agents or handoffs. See Multi-agent team and Sub-agent setup. After upgrade, smoke Workboard if you use board orchestration (upgrade runbook → v2026.6.1+).

• Review before apply — Even governed proposals can contain unsafe shell or network calls; read previews like any ClawHub skill.
• Exec approvals — Skills still load via the read tool under v2026.5.20+ policy (Exec approvals guide).
• Sandbox — Test new skills in a restricted profile before wide channel exposure (Security hardening).