Skill Workshop: Letting Claw Learn Skills on Its Own
Goal in one sentence: Enable and manage OpenClaw self-learning well—let Claw distill "learned skills" on its own after finishing complex work, and let you review, undo, archive, or turn it off entirely.
Prerequisite reading: Getting Started with Skills: Installation, Usage, and Configuration.
Preparation
- Version ≥ 2026.10.5: Workshop 2.0's version history and undo are available starting from this version; the current repository version is 2026.10.9.
skill-workshopis a built-in plugin, enabled by default, with no installation required.- First distinguish three homes: learned skills store "how-to"; facts about the user and the world go in the memory system; conventions and build commands for a single code repository go in that repository's docs or
AGENTS.md—Claw will not store them as learned skills.
Follow Along
Step 1: Confirm Self-Learning Is On
Action: Check that the configuration does not contain any of these three disabling items—plugins.entries.skill-workshop.enabled: false, plugins.enabled: false, or a plugins.allow allowlist that does not include skill-workshop. If any one of them takes effect, learned skills will not be loaded into Claw's prompt.
Expected: With a default installation, no changes are needed; Workshop is already running.
Step 2: Do a Hard Job Once and Let It Distill on Its Own
Action: Give Claw a multi-step task that will recur in the future. It has two learning paths. One is learning on the spot: if a learned skill leads the steps astray, it will review that skill and patch that step; after finishing a difficult multi-step job, it stores the workable approach as a new skill, or merges it into an existing skill covering that type of task. The other is background review: after the conversation accumulates 10 model iterations, a background run will review and save what is worth keeping; in propose mode, it first stores a draft and only finalizes it after you review it with /learn drafts. The self-learning mode is controlled by plugins.entries.skill-workshop.config.autonomous.mode, defaulting to auto (background review takes effect directly), with propose and off as options, for example openclaw config set plugins.entries.skill-workshop.config.autonomous.mode propose.
Expected: A skill written by Claw itself appears. Every change first saves the old version, takes effect immediately, and can be undone at any time; a running session still keeps the skill snapshot from when it started, and only new sessions load the changes.
Step 3: Review, Undo, Archive
Action: If you find a mistakenly learned skill, undo it to restore the previous version; to stop a skill from being used again, simply archive it; skills unused for 30 consecutive days are archived automatically.
Expected: After undoing, it returns to the old version; after archiving, the skill disappears from the prompt.
Verification
- Redo a similar task and see whether Claw follows the learned skill directly instead of groping step by step.
- Ask it to recite the steps of this skill to confirm the content matches expectations.
- Undo one change and confirm the old version has been restored.
Common Pitfalls
- Trying to block learned skills with a skill allowlist: Ineffective. Learned skills belong to a single agent and are always visible to it, bypassing the
agents.defaults.skillsandagents.entries.<id>.skillsallowlists; they can only be archived. - Treating content that belongs in memory or AGENTS.md as a skill: User preferences and facts go in memory; code repository conventions go in the repository's docs or
AGENTS.mdso that all agents working there can read them. - Looking for your own hand-written skills in Workshop: Workspace, project, ClawHub, plugin, and built-in skills are not Workshop skills; go back to the source to modify them, see Writing Your Own Skill.
- Version too old: Below 2026.10.5 there is no version history or undo; upgrade first.
- The official documentation also covers operational aspects such as
/learn [request](making Claw learn immediately from the current conversation or a specified source),/learn drafts(reviewing propose drafts),/learn undo <id>(undoing one review's changes, requires owner), and skill diet (slimming skills over 12 KB once a day), which this article does not expand on.
Next Lesson
Skills can now grow on their own; next, get permissions under control: Security Hardening (Part 1): Access Control and Identity.
Related reading: Writing Your Own Skill: SKILL.md and ClawHub Publishing, Memory System: Letting Claw Remember You.