# OpenClaw Skills A skill in [[OpenClaw]] is a directory containing a `SKILL.md` with YAML frontmatter and instructions that teach an agent how to use a tool or follow a procedure. The format is **AgentSkills-compatible**, the same shape used by [[Claude Code]] and other harnesses, which means most skills written for one runtime can be loaded by another with minimal changes. Skills are how OpenClaw stays composable: instead of cramming every behaviour into the core, the runtime stays small and capabilities arrive as discoverable, gateable, swappable directories. ## Discovery Order Loaded from three locations, **highest precedence first**: 1. Workspace ; `<workspace>/skills/` 2. Managed (per-user) ; `~/.openclaw/skills/` 3. Bundled (ships with the runtime) Workspace overrides managed overrides bundled. This is the same pattern OpenClaw uses elsewhere (hooks, agents) ; the override chain is consistent and predictable. ## Gating via Requirements A `SKILL.md` can declare requirements that must be satisfied before the skill loads: - **Required binaries** ; `git`, `ffmpeg`, `pandoc`, … - **Env vars** ; an API key must be present - **Config paths** ; specific config files must exist - **OS restrictions** ; macOS-only, Linux-only, Windows-only Skills that fail their requirements are silently skipped, not errored ; this is intentional so a workspace can ship aspirational skills without breaking on machines that don't have the prerequisites. ## ClawHub **ClawHub** ([clawhub.com](https://clawhub.com)) is the public skill registry. It plays the same role for OpenClaw that npm plays for Node, Homebrew for CLIs, or the Marketplace for [[GitHub Actions]]. ```bash clawhub install <skill-slug> clawhub update --all clawhub list --installed ``` The community-maintained list at [VoltAgent/awesome-openclaw-skills](https://github.com/VoltAgent/awesome-openclaw-skills) is a good index of what's worth installing. ## Self-Improving Loop OpenClaw's spec includes a *self-improving* property: an agent can author and modify its own skills. The loop is governed by [[OpenClaw Standing Orders|standing orders]] and the [[OpenClaw Hooks|hook system]] ; standing orders define the boundaries within which self-modification is allowed; hooks fire on key lifecycle events; new skills written into the workspace `skills/` directory take precedence on the next load. ## Plugin vs Skill - **Skills** teach an agent *how to do* something with existing tools - **Plugins** add *new* tools, channels, context engines, or hooks to the gateway itself A plugin is heavier (npm package + manifest + runtime entry point); a skill is lighter (a directory + YAML + Markdown). Pick the lightest abstraction that fits. ## References - Skills: https://docs.openclaw.ai/tools/skills - ClawHub: https://clawhub.com - AgentSkills spec: https://agentskills.io/specification.md ## Related - [[OpenClaw]] - [[OpenClaw Gateway]] - [[OpenClaw Hooks]] - [[OpenClaw Standing Orders]] - [[AI Agent Skills]] - [[Claude Code Skills]]