Module 5: Skills and Plugins

Extend agent capabilities with reusable components: skills for in-context instructions, plugins for external tools.

What You’ll Learn

1. Skills Overview — In-context instruction modules
2. Plugin Basics — External tool integrations
3. When to Use Which — Decision framework
4. Coach as Example — Complete skill case study

Time Required

~45 minutes for all sections

Skills vs Plugins

Both extend agent capabilities, but differently:

Aspect	Skills	Plugins
What	Markdown instructions	External code/services
Loaded	Into context at runtime	As callable tools
Language	Natural language	Code (JS, Python, etc.)
Execution	Agent interprets	Direct execution
Location	Workspace files	MCP servers, APIs

Skills

In-context instruction modules. Think “runbooks” or “playbooks”:

skills/
└── deployment/
    └── SKILL.md        # Instructions for deployment workflow

<!-- SKILL.md -->
# Deployment Skill

## Prerequisites
- All tests passing
- Branch up to date with main

## Steps
1. Create release branch
2. Update version in package.json
3. Generate changelog
4. Create PR
5. Wait for approval
6. Merge and tag

Agent reads SKILL.md and follows the instructions using its existing tools.

Plugins

External services that provide additional tools via MCP (Model Context Protocol):

// openclaw.json
{
  mcp: {
    servers: {
      "github": {
        command: "mcp-server-github",
        args: ["--token", "${GITHUB_TOKEN}"],
      },
    },
  },
}

Agent gains new tools: github_create_issue, github_list_prs, etc.

Quick Comparison

Task: “Create a GitHub issue”

With Skill (no plugin):

<!-- skills/github/SKILL.md -->
# GitHub Skill

## Create Issue
Use bash with gh CLI:
```bash
gh issue create --title "..." --body "..."

Agent reads skill, runs `bash` with `gh` command.

**With Plugin:**
```json5
// Direct tool call
{
  "tool": "github_create_issue",
  "params": {
    "title": "...",
    "body": "..."
  }
}

Agent calls plugin tool directly.

Trade-offs

Approach	Pros	Cons
Skill	No external deps, flexible	Uses context, may misinterpret
Plugin	Precise, fast, no context	Requires setup, less flexible

When to Use What

Need to extend agent capabilities?
│
├─ Does the task need external integration?
│  ├─ Yes → Plugin (MCP server)
│  └─ No → Probably a skill
│
├─ Is there complex logic involved?
│  ├─ Yes → Plugin (code handles logic)
│  └─ No → Skill works fine
│
├─ Do you need structured I/O?
│  ├─ Yes → Plugin (defined schemas)
│  └─ No → Skill (natural language)
│
└─ Will multiple agents use this?
   ├─ Yes → Consider either, but plugins are more consistent
   └─ No → Skill is simpler

Module Contents

Skills Overview

• Skill file structure
• How skills are loaded
• Writing effective skill instructions
• Skill libraries and organization

Plugin Basics

• MCP protocol overview
• Installing MCP servers
• Configuring plugins
• Available plugin ecosystem

When to Use Which

• Decision framework
• Hybrid approaches
• Migration patterns
• Performance considerations

Coach as Example

• Complete skill implementation
• SQLite database integration
• CLI tool design
• Cron-based scheduling
• GTD-style curation frontend

Next Steps

Start with Skills Overview to learn how to create in-context instruction modules →