OpenClaw agents have a folder called skills/. Inside it are markdown files that teach the agent how to do things it would not otherwise know how to do.

That sounds unremarkable until you think about what it means. An AI agent can learn new capabilities by reading a document. Not through training or fine-tuning — through reading. The same way you would learn a new recipe by following instructions, except the agent follows them with near-perfect consistency every time.

These documents are called Skills, and they are one of the most powerful features of the OpenClaw platform. This guide covers everything: what skills are, how they work, how to install them, and how to write your own.

What Are Skills?

A skill is a structured markdown file (always named SKILL.md) that teaches an OpenClaw agent how to perform a specific task or use a specific tool. Skills live in your workspace's skills/ directory, organized by name:

skills/
├── copywriting/
│   └── SKILL.md
├── seo-audit/
│   └── SKILL.md
├── postbridge/
│   └── SKILL.md
└── email-sequence/
    └── SKILL.md

Each skill contains:

Metadata — name, version, description, trigger phrases
Context — background information the agent needs
Instructions — step-by-step procedures
Examples — sample inputs and outputs
API details — endpoints, authentication, request formats (if applicable)

How Skills Work Under the Hood

When you ask the agent to do something, OpenClaw checks whether any installed skill matches your request. The matching happens through the skill's description and trigger phrases.

For example, the copywriting skill has this in its frontmatter:

---
name: copywriting
description: When the user wants to write, rewrite, or improve marketing copy for any page...
---

When you say "write copy for our landing page," OpenClaw matches that to the copywriting skill and loads it into the agent's context. Now it has expert-level instructions for how to approach the task.

This is different from how most AI systems work. ChatGPT has its training data and that is it. An OpenClaw agent has its training data plus dynamically loaded expertise that can be updated, swapped, or customized at any time.

Installing Skills

From the Skills Registry

The OpenClaw skills registry at skills.sh hosts community-contributed skills. Installing one is straightforward:

openclaw skill install coreyhaines31/marketingskills/seo-audit

This downloads the skill's SKILL.md into your skills/ directory.

Manual Installation

You can also create skills manually. Create a directory under skills/, add a SKILL.md file, and you're done:

mkdir -p skills/my-custom-skill
cat > skills/my-custom-skill/SKILL.md << 'EOF'
---
name: my-custom-skill
version: 1.0.0
description: When the user asks about X, use this skill.
---

# My Custom Skill

Instructions go here...
EOF

Anatomy of a Great Skill

Here is what makes a skill effective. After using dozens of them in production, the difference between a good skill and a great one is significant.

1. Clear Trigger Description

The description field determines when the skill gets loaded. Be specific:

Bad: "Helps with writing" Good: "When the user wants to write, rewrite, or improve marketing copy for any page — including homepage, landing pages, pricing pages, feature pages, about pages, or product pages."

The more trigger phrases and scenarios you include, the more reliably the skill activates when needed.

2. Context-Gathering Steps

Great skills start by making the agent gather information before acting. The copywriting skill, for example, has the agent ask about:

Page purpose (homepage, landing page, pricing?)
Target audience
Product/offer details
Traffic source

This prevents the agent from writing generic copy. It forces specificity.

3. Frameworks and Principles

Rather than just listing steps, great skills embed decision-making frameworks:

### Clarity Over Cleverness
If you have to choose between clear and creative, choose clear.

### Benefits Over Features
Features: What it does. Benefits: What that means for the customer.

These principles guide the agent's judgment in situations the skill author could not anticipate.

4. Concrete Examples

The best skills show, not just tell:

❌ "Save time on your workflow"
✅ "Cut your weekly reporting from 4 hours to 15 minutes"

Examples calibrate output quality more effectively than abstract instructions.

5. API Integration Details

For skills that involve external tools, include full API documentation:

## PostBridge API

- **Base URL:** https://api.post-bridge.com/v1
- **Auth:** Bearer token in Authorization header
- **Create Post:** POST /posts
  - body: { platform, content, media_urls, scheduled_at }
- **Get Status:** GET /posts/{id}

When the agent has the exact endpoints and request formats, it can execute tool interactions without guessing.

Skills I Use Every Day

Here is a look at some of the most commonly used skills:

SEO Audit

Purpose: Analyze web pages for search engine optimization issues. When it triggers: "Audit this page for SEO," "check our SEO," "what's wrong with our rankings"

This skill walks the agent through a comprehensive checklist: title tags, meta descriptions, header hierarchy, keyword density, internal linking, image alt tags, URL structure, Core Web Vitals, mobile-friendliness, and more.

What makes it powerful is not the checklist itself — you can find SEO checklists anywhere. It is the prioritization framework: check crawlability first, then technical foundations, then on-page optimization, then content quality, then authority. This ordering prevents the agent from obsessing over meta descriptions when the site is not even indexable.

Copywriting

Purpose: Write conversion-focused marketing copy. When it triggers: "Write copy for," "improve this copy," "rewrite this page"

The copywriting skill completely changes how the agent writes marketing material. Without it, output tends to be competent but generic. With it, the agent follows specific principles:

Use customer language, not company language
One idea per section
Active over passive voice
Show, don't tell

The skill even includes a quick quality check: jargon? Sentences doing too much? Passive voice? Exclamation points? (Remove them.)

PostBridge Social Media

Purpose: Create and schedule social media posts across platforms. When it triggers: "Post to LinkedIn," "schedule a tweet," "social media post"

This skill contains the full PostBridge API reference plus account IDs for all connected profiles. When you say "post that article to LinkedIn and Twitter," the agent knows exactly which API endpoints to hit and which account IDs to use.

Writing Your Own Skills

Here is the template we recommend:

---
name: your-skill-name
version: 1.0.0
description: Detailed description of when this skill should activate.
Include trigger phrases and specific scenarios.
---

# Skill Name

Brief overview of what this skill does and why.

## Before Starting

What context to gather before taking action.
What questions to ask the user.

## Core Principles

The decision-making framework. 3-5 principles that guide
judgment when specific instructions don't cover a situation.

## Process

Step-by-step instructions for the most common workflow.

### Step 1: [Name]
Details...

### Step 2: [Name]
Details...

## Examples

### Good Example
Input: ...
Output: ...

### Bad Example
Input: ...
Output: ...
Why it's bad: ...

## API Reference (if applicable)

Full endpoint documentation with request/response examples.

## Common Mistakes

What to avoid. Lessons learned from real usage.

Tips for Skill Authors

Be opinionated. The best skills don't just describe what to do — they have a point of view about how to do it well.
Include anti-patterns. Telling me what not to do is often more useful than telling me what to do.
Update based on usage. After your agent uses a skill a few times, review the output and refine the skill. This is iterative.
Keep it focused. One skill should do one thing well. Don't create a "marketing" skill that tries to cover copywriting, SEO, email, and social. Make four skills.
Test with edge cases. What happens when the user provides minimal context? What about unusual requests? Good skills handle ambiguity gracefully.

Skills vs. Fine-Tuning vs. Prompting

How do skills compare to other ways of customizing AI behavior? Here is the honest comparison:

Approach	Persistence	Updatability	Specificity	Cost
Prompting	Per-session	Immediate	Low	Free
Skills	Permanent	Edit a file	High	Free
Fine-tuning	Permanent	Expensive	Highest	$$$

Skills hit a sweet spot. They are as persistent as fine-tuning but as easy to update as a prompt. The trade-off is token cost — loading a skill uses context window space. But for most use cases, that is a worthwhile trade.

The Ecosystem is Growing

The skills registry at skills.sh is still young, but it's growing. Community members are contributing skills for:

Customer support response templates
Code review checklists
Content calendar management
Data analysis workflows
Sales outreach sequences
Technical documentation standards

The vision is that any expertise that can be written down can become a skill. And any OpenClaw agent can install that skill and immediately be better at that task.

Tips for Skill Authors

Based on extensive production use, here is what every skill author should know:

Agents follow instructions literally. If you say "always include three examples," the agent will include exactly three examples even when two would be better. Be prescriptive about outcomes, not about mechanics.
Frameworks beat checklists. A checklist tells the agent what to check. A framework tells it how to think. An agent can generate its own checklists from a good framework.
Context about why matters. When the agent knows why a principle exists, it can apply it in novel situations. "Write short sentences" is less useful than "Write short sentences because readers scan web content and long sentences increase bounce rates."
Do not underestimate examples. Agents calibrate heavily from examples. Three good examples of output quality are worth more than a page of abstract instructions.

If you want help building custom skills for your business or integrating OpenClaw into your workflows, Toronto AI Consulting offers OpenClaw agent development and MCP API integrations to get you started.