Documentation
Agent Tuning
Skills

Skills

Skills let you turn a repeatable workflow into something consistent: structure, checklists, tool sequences, and “how to behave” guidelines that the agent can reuse across tasks.

Start small (easier to iterate)

  • Codify the output structure first (headings/fields/order).
  • Keep skills short and testable (one job per skill).
  • Add tool usage gradually (don’t start with a huge “do everything” skill).

What a skill looks like

A skill is usually a directory containing at least SKILL.md:

~/.openclaw/skills/<skill-name>/SKILL.md

You can also put project-specific skills in your workspace (higher priority):

<workspace>/skills/<skill-name>/SKILL.md

In SKILL.md, write:

  • when to use it
  • steps/checklist (including tool calls if needed)
  • expected output format
  • safety constraints (what to avoid)

Where skills load from (precedence)

OpenClaw commonly uses two layers:

  1. Global skills: ~/.openclaw/skills/
  2. Workspace skills: <workspace>/skills/ (wins over global)

This lets you keep general skills globally, and override or specialize per project.

Debugging and self-checks 

openclaw skills list
openclaw skills info <name>
openclaw skills check

If you update a skill and don’t see it taking effect, restart the gateway:

openclaw gateway restart

Security note

Skills influence the agent’s behavior and capability surface. Before installing a third‑party skill, read SKILL.md and make sure you trust the source.

Next links

Multi-AgentWorkspace Files