Building Plugins for Claude Code

Expert guidance for creating well-structured plugins, skills, hooks, agents, commands, and MCP configurations.

Core Principles (Anthropic Official)

1. Composable - Multiple skills work together seamlessly, Claude coordinates automatically
2. Portable - Same format works across Claude apps, Claude Code, and Claude API
3. Efficient - Progressive disclosure loads only what's needed (description → SKILL.md → references)
4. Powerful - Combines prompts + executable code for complete workflows
5. Secure - Never hardcode secrets, use ${CLAUDE_PLUGIN_ROOT} for paths

Critical Budget Limits

Plugin Structure

my-plugin/
├── plugin.json              # Manifest (at root for marketplace plugins)
├── commands/                # Slash commands (.md files)
├── skills/                  # Skills with SKILL.md
│   └── skill-name/
│       ├── SKILL.md
│       └── references/
├── agents/                  # Sub-agents (.md files)
├── hooks/
│   └── hooks.json           # Event handlers
├── .mcp.json                # MCP servers
└── README.md

Quick Reference

plugin.json (Minimum)

{
  "name": "plugin-name",
  "version": "1.0.0",
  "description": "What this plugin does"
}

SKILL.md Frontmatter

---
name: skill-name
description: Use when [trigger]. For [use cases].
allowed-tools: Read, Glob, Grep
model: sonnet
context: fork
user-invocable: true
---

Agent Frontmatter

---
name: agent-name
description: Expert in [domain]. Use for [tasks].
tools: [Read, Grep, Glob]
disallowedTools: [Write]
model: sonnet
permissionMode: default
color: blue
---

Command Frontmatter

---
description: What this command does
allowed-tools: Read, Write, Edit
argument-hint: "<required> [optional]"
---

Hook Configuration

{
  "hooks": {
    "PreToolUse": [{
      "matcher": "Write|Edit",
      "hooks": [{ "type": "command", "command": "${CLAUDE_PLUGIN_ROOT}/scripts/validate.sh" }]
    }]
  }
}

Detailed References

Load these for comprehensive documentation:

Common Tasks

Creating a New Plugin

1. Create directory with plugin.json at root
2. Add skills in skills/skill-name/SKILL.md
3. Add commands in commands/command-name.md
4. Test with claude --plugin-dir ./my-plugin
5. Validate with claude plugin validate .

Creating a Skill

1. Create skills/my-skill/SKILL.md with frontmatter
2. Write concise instructions (<500 lines)
3. Add references/ for detailed documentation
4. Test by invoking /my-skill

Adding Hooks

1. Create hooks/hooks.json
2. Define event matchers and hook types
3. Restart Claude Code (hooks load at session start)

Creating a Marketplace

1. Create .claude-plugin/marketplace.json at repo root
2. Add plugins with plugin.json at each plugin root
3. Test: /plugin marketplace add ./my-marketplace

Development Methodology (Anthropic Engineering)

Evaluation-First Approach:

1. Run representative tasks WITHOUT skills first
2. Observe where Claude struggles or requires additional context
3. Build skills incrementally to address specific gaps
4. Use Claude-assisted skill authoring (ask Claude to draft skills after successful workflows)

Self-Reflection Pattern: After completing a task successfully, ask Claude:

"What worked well? Draft a SKILL.md that captures this workflow for next time."

Best Practices Checklist

• Plugin name is lowercase with hyphens, 3-64 characters
• Description includes trigger phrases for auto-discovery
• Description under 400 chars (not 1024 max) to preserve budget
• Description uses third person voice
• Secrets use environment variables, never hardcoded
• Main files are concise, details in reference files
• Uses ${CLAUDE_PLUGIN_ROOT} for all internal paths
• Semantic versioning (MAJOR.MINOR.PATCH)
• README with examples
• Tested with Haiku, Sonnet, and Opus
• Skills trigger on expected phrases (test auto-activation)