#OpenClaw Skills

Skills package multi-step flows into reusable, shareable, portable units of capability. OpenClaw's skills are plain-text files — transparent and easy to version.

#What Is a Skill?

A skill is a reusable procedure described in natural language, stored as a SKILL.md file with YAML frontmatter (metadata):

---
name: weekly-report
description: Summarize this week's GitHub activity into a report
---

# Steps
1. Fetch this week's commits and issues via browser/API
2. Distill into no more than 5 bullet points
3. Generate a Markdown report and send it to Slack

• Reusable: invoke it next time without re-deriving the steps
• Portable: format compatible with Claude Code / Cursor conventions
• Shareable: distributed via the ClawHub skill registry, community repos, or direct URLs
• Auto-draftable: when a skill is missing, describe the task and have the agent draft one

#Three Ways to Get Skills

⚠️ Security first: ClawHub had 230+ malicious skills uploaded in early February 2026, and analysis found ~26% of community skills contained at least one vulnerability. Skills are not sandboxed by default — review the contents of any third-party skill before installing. See Security.

#Where Skills Live

Skills are stored as SKILL.md files globally or per workspace:

~/.openclaw/skills/        # Global skills
<workspace>/skills/        # Workspace-level skills
<workspace>/SKILL.md       # Single-skill file

Because they're plain text, you can version them with Git, review them in an editor, and grep them — fully transparent.

#Auto-Drafting Skills

A highlight of OpenClaw: when no skill exists, let the agent write one.

You: Make me a skill that summarizes unread email each morning and sends it to Telegram
OpenClaw: drafts a SKILL.md (with steps) → you review/tweak → save and reuse

This way your skill library gradually fits your own workflow.

#Writing Your Own Skills

When hand-writing skills:

• Clear frontmatter: name and description determine whether retrieval hits it
• Single responsibility: one skill does one thing, easy to compose
• Spell out steps and required tools, flag high-risk actions (for approval)
• Verify once on a low-risk scenario before putting it into use

#Practical Tips

• Prioritize distilling things you ask OpenClaw to do repeatedly into skills
• Read community skills line by line before installing, especially shell / network / secret parts
• Review even the skills the agent drafts — don't let it assume
• Manage skills/ with Git so teams can share and code-review them

#Next Steps

• Memory System — skills handle "doing," memory handles "remembering"
• Permissions & Security — skill security is paramount
• Multi-Agent — reuse skills across sessions