•Lucas Lyu

Building Capabilities: How to Write Skills for OpenClaw

As I build out my local AI stack with OpenClaw, one of the most powerful features I've encountered is the Skill System.

Unlike many agent frameworks that rely on brittle API integrations or complex Python bindings, OpenClaw takes a Unix-philosophy approach: if it has a CLI, it can be a skill.

What is a Skill?

In OpenClaw, a skill is essentially a definition file (SKILL.md) that teaches the AI two things:

  1. When to use a specific tool (semantic description).
  2. How to invoke it (CLI syntax).

It's a bridge between natural language intent ("Check the weather") and machine execution (curl wttr.in).

The Anatomy of a Skill

Here is the structure I use when creating a new capability. Let's say we want to give the agent the ability to manage Docker containers.

We create a directory skills/docker/ and add a SKILL.md:

# Docker Skill

Description: Manage Docker containers and images. Use when the user asks to list running containers, inspect logs, or restart services.

## Tools

### docker_ps
- Description: List running containers.
- Usage: `docker ps --format "table {{.ID}}\t{{.Image}}\t{{.Status}}\t{{.Names}}"`

### docker_logs
- Description: Fetch logs for a specific container.
- Usage: `docker logs --tail 50 <container_id>`

Why This Matters

This approach has massive advantages for Personal AI:

  1. Security: You define exactly what flags are allowed. No blind rm -rf.
  2. Portability: It works with any binary on your system (kubectl, gh, ffmpeg).
  3. Low Code: You don't need to write a wrapper library. You just write Markdown.

My Workflow

I'm currently maintaining a collection of private skills for my daily workflows, including:

  • University Ops: Fetching canvas assignments via API.
  • Home Lab: Managing my Plex server and Nginx proxy.
  • Coding: Auto-generating extensive test suites.

By treating "Agent Capabilities" as configuration rather than code, I can iterate much faster. The agent reads the manual (the SKILL.md) and instantly knows how to use the tool.

Stay tuned for my next post where I'll share my "University Ops" skill configuration.