OpenClaw has shipped Skill Workshop, a feature designed to solve a real problem with AI agents: turning one-off tasks into reusable behaviors without the risk of locking in bad patterns permanently.

The Proposal-First Approach

When an agent creates or revises a skill through Skill Workshop, it starts as a proposal rather than immediately modifying live skills. During this pending state, the file exists as PROPOSAL.md and does not influence agent behavior yet. This gives you time to review, tweak, or reject before anything becomes baked into future runs. The loop mirrors normal collaboration: you ask for a reusable workflow, the agent drafts a proposed skill, you provide feedback, and the agent revises. Only after your approval does it become active as SKILL.md. No manual file creation required, no guessing where skills should live across chat, UI, or CLI contexts.

Two Views for Different Workflows

The Control UI offers Board view as the full workshop experience. You can move through pending, applied, rejected, and stale proposals while searching, inspecting, previews of support files, and seeing exactly what changed in each revision. Today view provides a faster pass optimized for moving through proposed skills one at a time. It presents the next proposal and asks a concrete question: should this become part of your skill set? Use it when ready, tweak when close, or skip entirely when it should not persist. The Tweak option is where Skill Workshop matters mostβ€”generated work is often almost right but needs wording adjustments, missing steps, safer fallbacks, or file type changes like converting a script to a template.

Support Files Travel With Proposals

Useful skills rarely exist in isolation. A digest skill might need a response template, a debugging skill may require example logs, and a deployment workflow could depend on smoke-test scripts. Skill Workshop keeps these supporting files bundled with the proposal through folders labeled assets, examples, references, scripts, and templates. Support files are shown in the UI before you apply them, scanned alongside the proposal, and written to disk only when you approve the skill. The system deliberately restricts pathsβ€”no absolute paths, no traversal attacks, no hidden segments, and no writes outside the skill directory itself.

Key Takeaways

  • Skills start as proposals (PROPOSAL.md) until explicitly approved or rejected
  • Board view offers full workshop controls; Today view accelerates single-skill reviews
  • Supporting files travel with proposals across all access points: chat, UI, channels, CLI, and Gateway
  • Path restrictions prevent skills from writing outside their intended boundaries

The Bottom Line

Skill Workshop fills a gap that most agent frameworks ignore entirely. Creating reusable behavior without a review step is how you accidentally propagate broken patterns across your entire workflow. OpenClaw's approach makes the feedback loop explicit, which is exactly what builders need when handing off repeatable tasks to AI systems.