Skills

Use OpenClaw skills to teach repeatable workflows, review rubrics, operating rules, and service-specific behavior without writing runtime code.

Skills are Markdown instruction packs. They teach the agent how to use existing capabilities. They do not add a new runtime by themselves.

Use a skill when you can describe the workflow in words and the required tools already exist.

Skill Anatomy

A skill is a directory with a SKILL.md file:

Minimal example:

---
name: incident-summary
description: Summarize incident threads and produce a concise action list
---

## When to use

Use this skill when the user asks for an incident recap.

## Workflow

1. Read the provided thread or transcript.
2. Identify timeline, impact, root cause, and follow-ups.
3. Produce a short summary and owner-tagged action list.

Use {baseDir} to reference files bundled with this skill.

Loading Order

OpenClaw loads skills from several roots, highest precedence first:

When duplicates exist, higher-priority locations win. This lets you override a bundled, managed, or plugin-packaged skill for one workspace without changing the global copy.

Skill roots can contain grouped folders. The skill name, slash command, and allowlist key come from the name frontmatter field, or from the directory name when name is missing.

Per-Agent Skill Control

In multi-agent setups, skills can be scoped per agent. Use a shared baseline for common behavior and per-agent allowlists when one persona needs a smaller or specialized skill set.

Examples:

• a personal agent gets calendar and inbox workflow skills
• a coding agent gets review, test, and release skills
• a support agent gets escalation and CRM skills

Allowlist details matter:

• omit agents.defaults.skills to leave skills unrestricted by default
• omit agents.list[].skills to inherit the default allowlist
• set agents.list[].skills: [] to expose no skills for that agent
• set a non-empty agents.list[].skills list to replace defaults entirely, not merge with them

ClawHub

ClawHub is the discovery and publishing surface for OpenClaw skills and plugins. Use it to find, install, update, audit, and publish packages.

openclaw skills search calendar
openclaw skills install <skill-name>
openclaw skills list
openclaw skills update --all
openclaw skills verify <skill-name>

Prefer audited, narrow skills over broad instruction packs that ask for excessive tool access.

Skill Workshop

Skill Workshop helps create or update workspace skills through a guided review loop. It is useful when you want the agent to draft an instruction pack, then inspect and approve the result before use.

Use it for:

• turning a repeated prompt into a reusable skill
• documenting a team workflow
• adding service-specific checklists
• refining a review rubric

Requirements and Gating

Skills can declare requirements so they activate only when prerequisites are available:

---
name: github-release-helper
description: Prepare GitHub release notes and validation checklists
metadata:
  openclaw:
    requires:
      bins: ["gh"]
      env: ["GITHUB_TOKEN"]
---

Common gates include binaries, environment variables, and config paths.

skills.entries.<key>.env and skills.entries.<key>.apiKey can inject secrets into the host agent run for that turn, but they do not automatically enter the sandbox. Keep secrets out of skill bodies and prompt text.

Security Review

Read a skill before enabling it. A skill can change the agent's system prompt and influence tool use.

Review for:

• hidden credential requests
• instructions to bypass approval
• broad file or network access assumptions
• unclear external side effects
• stale command names

Use sandboxing and narrow tool profiles for third-party or experimental skills.

Token Cost

Every active skill adds prompt text. Keep descriptions short and load only the skills relevant to the current agent. Large reference files should be opened on demand rather than always injected.