Skip to main content
Skills turn the way you work into reusable instructions your Claw can follow again, and this release connects the entire path. You can create and validate a skill, find or install it, call it directly from a conversation, review proposed changes, and have supported edits ready on the next turn of a persistent session. Invalid skills are reported individually, so the rest of the catalog remains available. Skill Workshop brings proposals, checks, decisions, and applied history into one workflow. Self-learning can turn substantial work and durable corrections into proposed improvements, or maintain skills created through Workshop when automatic learning is enabled, while skills owned by you or someone else stay under their owner’s control.
Creating a skill now follows one guided path from choosing how it should be invoked through adding supporting files, saving it, and validating the result. The checker understands supported invocation metadata and catches problems such as an overlong description before anything is written.OpenClaw now reports malformed metadata, unreadable files, oversized instructions, and shadowed copies against the skill that caused them, while continuing to load valid skills around it. In a persistent Gateway session, edits to canonical and managed-worktree skills are available on the next turn, and required skill instructions are read in full.
Installed skills, ClawHub discovery, skill settings, and Skill Workshop now share one Plugins hub, giving you one place to find a skill, install it, configure it, and confirm its current status. Skills and plugins keep their separate lifecycles, while reconnecting or switching the active agent, model, or connectors refreshes the lists from the current Gateway.When a skill is available to you and the active agent, you can choose it in chat or name up to eight with $skill-name across supported chat and agent entry points, including eligible skills hidden from automatic model selection. The chat picker adds references to your draft without sending it, and Code Mode can list and read eligible skills within its existing sandbox and allowlist. Large model-visible catalogs can still be compacted, so openclaw skills check remains the complete inventory.
Skill Workshop gives you one place to turn an idea or a reusable lesson from substantial past work into a reviewable skill change. You can inspect the proposed instructions and supporting files, see results from plugin-provided scanners, benchmarks, and graders, revise the proposal, and then apply, reject, or quarantine it. Past-work scans produce pending proposals rather than editing live skills, and Android users can search and inspect them before an authenticated administrator makes a change.Every decision stays bound to the exact proposal revision you reviewed, so a later revision returns for review. Critical prompt-injection findings block application, interrupted applies can recover without overwriting a target changed elsewhere, and an explicitly selected remote Gateway remains the authority for the change. Applied revisions are grouped by skill with newest-first history and comparisons that say when the visible diff is incomplete.
OpenClaw can turn substantial work and durable corrections into reusable skills, then improve the Workshop-created skills that actually shaped a run. New and unconfigured installations start in auto, while upgrades keep their existing choice. off disables automatic repair, propose queues changes for review, and auto can create or update Workshop-owned skills with targeted patches or a same-turn repair. The conversation already in progress keeps the version it loaded until the next turn.Skills you wrote and shared skills owned elsewhere remain yours. Automatic learning can suggest improvements to them, but it cannot rewrite or remove them on its own, and explicit /learn or past-work scans also produce proposals for review.On supported agent runtimes, optional background review runs separately without interrupting or posting into chat. When both the learning mode and scheduled-job settings allow it, a visible weekly job reviews the collection, records usage and outcomes, preserves specialized skills, and creates recoverable backups. Restoring a backup remains an explicit choice.