Do I need programming skills for skills?

Not for pure prompt-based skills. Yes for complex tool chains or custom MCP servers. Markdown + YAML covers 80 % of cases.

Can a skill call multiple LLMs?

Yes. Via sub-skills or explicit provider switches inside a skill. More complex but powerful.

Yes. Frontmatter field version: 1.2.0 . On calver updates of OpenClaw, skills are not overwritten — they live in your team repo.

Guide

Create OpenClaw custom skills yourself

Standard skills are nice, but the most productive ones are tailored to your workflow. Full guide to building your own.

Manuel Streit

/ May 11, 2026 / 3 min read

About this guide

Create your own OpenClaw skills: skill file format, system prompt, tool calls, tests, deploy via Git and best practices.

Skill format

Skills are Markdown files with YAML frontmatter, in ~/.openclaw/skills/ or in your team repo. Example:

---
name: weekly-report
description: Generates a weekly report from CRM activity
trigger: ["/report", "weekly report"]
provider: anthropic
model: claude-sonnet-4-6
tools: [crm-read, calendar-read]
---

You are a report generator. Your job is to produce a short
weekly report from the last 7 days of CRM activity.

Format: Markdown, max 500 words.
Sections: Wins, Lost Deals, Pending, Next Week.

Define tool calls

Tools are MCP servers the skill may use. tools: [crm-read, calendar-read] allows only these two. Write access (crm-write) must be granted explicitly.

Testing

Per skill a test set. ~/.openclaw/skills/weekly-report.tests.json with example inputs and expected outputs. openclaw skill test weekly-report runs them.

Deploy via Git

Skills in a private Git repo, pull request for every change. openclaw skill sync pulls latest skills. In team setups: a central master daemon syncs, sub-daemons fetch.

Best practices

One skill = one clearly defined goal (no kitchen sink)
Set token limit per call
Embed examples in the system prompt ("few-shot")
Tests against real data, not synthetic
Version numbers in frontmatter (for rollback)