Building Plugins Skill

You are an expert at creating Claude Code plugins. Plugins are bundled packages that combine agents, skills, commands, and hooks into cohesive, distributable units.

What is a Plugin?

A plugin is a package that bundles related Claude Code components:

• Agents: Specialized subagents for delegated tasks
• Skills: Auto-invoked expertise modules
• Commands: User-triggered slash commands
• Hooks: Event-driven automation

Plugins enable users to install complete functionality with a single command.

When to Create a Plugin vs Individual Components

Use a PLUGIN when:

• You want to distribute multiple related components together
• You're building a cohesive feature set or domain expertise
• You want users to install everything with one command
• You need to maintain version compatibility across components
• You're creating a reusable toolkit for a specific domain

Use INDIVIDUAL COMPONENTS when:

• You only need a single agent, skill, or command
• Components are unrelated and can be used independently
• You're customizing for a specific project
• You don't plan to distribute or share

Plugin Creation Process

Creating a plugin involves:

1. Gathering requirements (name, components, metadata)
2. Creating the directory structure
3. Writing the plugin.json manifest
4. Creating each component (agents, skills, commands, hooks)
5. Writing comprehensive README.md
6. Validating the complete plugin

Component Creation: Each component type (agents, skills, commands, hooks) should follow its respective best practices. Use the corresponding building-* skills for expertise on creating each type.

Plugin Structure & Schema

Directory Structure

plugin-name/
├── .claude-plugin/
│   └── plugin.json          # Required: Plugin manifest
├── agents/                  # Optional: Agent definitions
│   ├── agent1.md
│   └── agent2.md
├── skills/                  # Optional: Skill directories
│   ├── skill1/
│   │   ├── SKILL.md
│   │   ├── scripts/
│   │   ├── references/
│   │   └── assets/
│   └── skill2/
│       └── SKILL.md
├── commands/                # Optional: Slash commands
│   ├── command1.md
│   └── command2.md
├── hooks/                   # Optional: Event hooks
│   ├── hooks.json
│   └── scripts/
├── scripts/                 # Optional: Helper scripts
│   └── setup.sh
├── .mcp.json               # Optional: MCP server configuration
└── README.md               # Required: Documentation

Minimal Plugin Structure

The absolute minimum for a valid plugin:

my-plugin/
├── .claude-plugin/
│   └── plugin.json
└── README.md

plugin.json Schema

Required Fields

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

Recommended Fields

{
  "name": "plugin-name",
  "version": "1.0.0",
  "description": "Comprehensive description of plugin functionality",
  "author": {
    "name": "Your Name",
    "email": "your.email@example.com",
    "url": "https://github.com/yourname"
  },
  "homepage": "https://github.com/yourname/plugin-name",
  "repository": "https://github.com/yourname/plugin-name",
  "license": "MIT",
  "keywords": ["keyword1", "keyword2", "keyword3"]
}

Component Registration

{
  "commands": "./commands/",
  "agents": ["./agents/agent1.md", "./agents/agent2.md"],
  "skills": "./skills/",
  "hooks": ["./hooks/hooks.json"]
}

Notes:

• Use directory paths ("./commands/") to include all files in a directory
• Use file arrays (["file1.md", "file2.md"]) to list specific files
• Paths are relative to plugin root directory

⚠️ CRITICAL FORMAT WARNING

Arrays MUST contain simple path strings, NOT objects!

❌ WRONG (will silently fail to load):

"commands": [
  {"name": "init", "path": "./commands/init.md", "description": "..."},
  {"name": "status", "path": "./commands/status.md"}
]

✅ CORRECT:

"commands": [
  "./commands/init.md",
  "./commands/status.md"
]

This applies to all component arrays: agents, skills, commands, and hooks.

Also note: Single-item arrays must still be arrays, not strings:

• ❌ "agents": "./agents/my-agent.md" (string - won't load)
• ✅ "agents": ["./agents/my-agent.md"] (array - correct)

Optional: MCP Servers

{
  "mcpServers": {
    "server-name": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-name"],
      "env": {
        "API_KEY": "${API_KEY}"
      }
    }
  }
}

Naming Conventions

Plugin Name:

• Lowercase letters, numbers, and hyphens only (no underscores!)
• Max 64 characters
• Descriptive and domain-specific
• Examples: code-review-suite, data-analytics-tools, git-workflow-automation

Component Names:

• Follow individual component naming rules
• Agents: Action-oriented (code-reviewer, test-generator)
• Skills: Gerund form preferred (analyzing-data, reviewing-code)
• Commands: Verb-first (new-feature, run-tests)
• Consistency: Use similar naming patterns within a plugin

Semantic Versioning

Plugins must follow semantic versioning: MAJOR.MINOR.PATCH

• MAJOR: Breaking changes (e.g., removed components, changed interfaces)
• MINOR: New features (e.g., new commands, enhanced capabilities)
• PATCH: Bug fixes and minor improvements

Examples:

• 1.0.0 → Initial release
• 1.1.0 → Added new command
• 1.1.1 → Fixed bug in existing command
• 2.0.0 → Removed deprecated agent (breaking change)

Creating a Plugin

Follow these steps to create a well-structured plugin:

Step 1: Gather Requirements

Ask the user:

1. Plugin name and purpose: What will this plugin do?
2. Target domain: What problem does it solve?
3. Components needed:
   • How many agents? What tasks?
   • How many skills? What expertise?
   • How many commands? What workflows?
   • Any hooks? What events?
4. Metadata:
   • Author information
   • License type (MIT, Apache 2.0, etc.)
   • Repository URL
   • Keywords for searchability

Step 2: Design Plugin Architecture

Plan the component structure:

Example: Code Review Plugin

code-review-suite/
├── agents/
│   ├── code-reviewer.md          # Deep code analysis
│   └── security-auditor.md       # Security scanning
├── skills/
│   ├── reviewing-code/           # Always-on review expertise
│   └── detecting-vulnerabilities/ # Security pattern matching
├── commands/
│   ├── review.md                 # /review [file]
│   ├── security-scan.md          # /security-scan
│   └── suggest-improvements.md   # /suggest-improvements
└── hooks/
    └── hooks.json                # Pre-commit validation

Design Principles:

• Cohesion: Components should work together toward a common goal
• Single Responsibility: Each component has a clear, focused purpose
• Minimal Overlap: Avoid duplicating functionality
• Progressive Complexity: Start simple, add features iteratively

Step 3: Create Directory Structure

mkdir -p plugin-name/.claude-plugin
mkdir -p plugin-name/agents
mkdir -p plugin-name/skills
mkdir -p plugin-name/commands
mkdir -p plugin-name/hooks
mkdir -p plugin-name/scripts

Step 4: Create plugin.json Manifest

Use the plugin.json schema template and populate all fields:

{
  "name": "plugin-name",
  "version": "1.0.0",
  "description": "Detailed description of what this plugin provides",
  "author": {
    "name": "Author Name",
    "email": "email@example.com",
    "url": "https://github.com/username"
  },
  "homepage": "https://github.com/username/plugin-name",
  "repository": "https://github.com/username/plugin-name",
  "license": "MIT",
  "keywords": ["domain", "automation", "tools"],
  "commands": "./commands/",
  "agents": "./agents/",
  "skills": "./skills/",
  "hooks": ["./hooks/hooks.json"]
}

Critical Validation:

• Valid JSON syntax (use python3 -m json.tool plugin.json)
• Name follows lowercase-hyphens convention
• Version is semantic (X.Y.Z)
• All paths reference actual directories/files

Step 5: Create Components

Create each component using the appropriate expertise:

For Agents:

• Follow the building-agents skill guidance
• Use the agent-template.md as a starting point
• Place in plugin-name/agents/

For Skills:

• Follow the building-skills skill guidance
• Create a directory with SKILL.md
• Place in plugin-name/skills/skill-name/

For Commands:

• Follow the building-commands skill guidance
• Use command-template.md as a starting point
• Place in plugin-name/commands/

For Hooks:

• Follow the building-hooks skill guidance
• Create hooks.json configuration
• Place in plugin-name/hooks/

Step 6: Write Comprehensive README.md

Use the README template from {baseDir}/templates/plugin-readme-template.md.

Required Sections:

1. Title and Description: What the plugin does
2. Features: Key capabilities
3. Installation: How to install
4. Components: List all agents/skills/commands/hooks
5. Usage: Examples and workflows
6. Configuration: Any setup required
7. License: License information

Optional Sections:

• Screenshots/demos
• Architecture diagrams
• Troubleshooting
• Contributing guidelines
• Changelog

Step 7: Validate the Plugin

Run the validation script:

python3 {baseDir}/scripts/validate-plugin.py plugin-name/

Validation Checks:

• [ ] plugin.json exists and has valid JSON
• [ ] Required fields present (name, version, description)
• [ ] Name follows conventions (lowercase-hyphens, max 64 chars)
• [ ] Version follows semantic versioning
• [ ] All referenced paths exist
• [ ] All components are valid (agents, skills, commands, hooks)
• [ ] README.md exists and is comprehensive
• [ ] License file exists (if license specified)
• [ ] No security issues (exposed secrets, dangerous scripts)

Step 8: Test the Plugin

Testing Checklist:

1. Installation Test: Symlink to .claude/plugins/ and verify Claude loads it
2. Component Tests:
   • Invoke each agent manually
   • Trigger skill auto-invocation
   • Run each command with various arguments
   • Trigger hooks with relevant events
3. Integration Tests: Verify components work together
4. Edge Cases: Test with invalid inputs, missing files, etc.

Step 9: Document Usage

Provide clear instructions:

## Installation

### Manual Installation
1. Clone this repository
2. Symlink to Claude's plugin directory:
   ```bash
   ln -s /path/to/plugin-name ~/.claude/plugins/plugin-name

3. Restart Claude Code

Marketplace Installation (if published)

claude plugin install plugin-name

Quick Start

1. Run your first command:

   /plugin-name:command arg1 arg2

2. Invoke an agent:

   Ask Claude to use the agent-name agent

3. Auto-invoked skills: Skills activate automatically when relevant.

Plugin Templates

This skill provides three plugin templates for different use cases:

1. Minimal Plugin Template

File: {baseDir}/templates/minimal-plugin-template/

Use when:

• Creating a simple, single-purpose plugin
• Only need 1-2 components
• Minimal complexity

Structure:

minimal-plugin/
├── .claude-plugin/plugin.json
├── commands/
│   └── main-command.md
└── README.md

2. Standard Plugin Template

File: {baseDir}/templates/standard-plugin-template/

Use when:

• Building a typical plugin with multiple components
• Need agents + commands or skills + hooks
• Moderate complexity

Structure:

standard-plugin/
├── .claude-plugin/plugin.json
├── agents/
│   └── main-agent.md
├── commands/
│   ├── command1.md
│   └── command2.md
├── scripts/
│   └── helper.sh
└── README.md

3. Full Plugin Template

File: {baseDir}/templates/full-plugin-template/

Use when:

• Building a comprehensive plugin suite
• Need all component types
• High complexity with multiple integrations

Structure:

full-plugin/
├── .claude-plugin/plugin.json
├── agents/
│   ├── agent1.md
│   └── agent2.md
├── skills/
│   ├── skill1/
│   │   ├── SKILL.md
│   │   └── scripts/
│   └── skill2/
│       └── SKILL.md
├── commands/
│   ├── cmd1.md
│   ├── cmd2.md
│   └── cmd3.md
├── hooks/
│   ├── hooks.json
│   └── scripts/
├── scripts/
│   └── setup.sh
├── .mcp.json
├── LICENSE
└── README.md

Common Plugin Patterns

Pattern 1: Development Tools Plugin

Purpose: Automate common development workflows

Components:

• Agents: code-reviewer, test-generator, refactoring-assistant
• Skills: reviewing-code, writing-tests, refactoring-code
• Commands: /format, /lint, /test, /build
• Hooks: PreToolUse for code quality checks

Example: dev-tools-suite, code-quality-automation

Pattern 2: Domain Expertise Plugin

Purpose: Provide specialized knowledge for a domain

Components:

• Skills: Domain-specific expertise (auto-invoked)
• Commands: Workflows specific to the domain
• Agents: Deep analysis for complex domain tasks

Example: data-analytics-tools, api-design-suite, security-analysis

Pattern 3: Workflow Automation Plugin

Purpose: Automate repetitive tasks and processes

Components:

• Commands: User-triggered workflows
• Hooks: Event-driven automation
• Scripts: Helper utilities
• Skills: Background expertise for automation

Example: git-workflow-automation, deployment-automation, project-scaffolding

Pattern 4: Integration Plugin

Purpose: Connect Claude to external tools and services

Components:

• MCP Servers: External service connections
• Commands: Trigger integrations
• Agents: Process external data
• Skills: Context about external services

Example: github-integration, jira-connector, database-tools

Marketplace Integration

If you're creating plugins for the Claude Code marketplace repository, you MUST maintain the central registry.

marketplace.json Registration

File: .claude-plugin/marketplace.json (at repository root)

This file is the central registry for all plugins in the marketplace.

When Adding a NEW Plugin

Update .claude-plugin/marketplace.json:

{
  "metadata": {
    "name": "Claude Code Plugin Marketplace",
    "version": "X.Y.Z",  // ← Increment MINOR version
    "stats": {
      "totalPlugins": N,  // ← Increment count
      "lastUpdated": "YYYY-MM-DD"  // ← Update date
    }
  },
  "plugins": [
    // ... existing plugins ...
    {
      "name": "new-plugin-name",
      "source": "./new-plugin-name",  // ← Path to plugin directory
      "description": "Plugin description",
      "version": "1.0.0",
      "category": "development-tools",  // or "automation", "integration", etc.
      "keywords": ["keyword1", "keyword2"],
      "author": {
        "name": "Author Name",
        "url": "https://github.com/username"
      },
      "repository": "https://github.com/username/repo",
      "license": "MIT",
      "homepage": "https://github.com/username/repo/tree/main/plugin-name"
    }
  ]
}

When Updating an EXISTING Plugin

Update both files:

1. Plugin's plugin.json:

• Increment version (following semantic versioning)
• Update description if changed
• Update components array if changed

2. Root marketplace.json:

{
  "metadata": {
    "version": "X.Y.Z",  // ← Increment PATCH version
    "stats": {
      "lastUpdated": "YYYY-MM-DD"  // ← Update date
    }
  },
  "plugins": [
    {
      "name": "existing-plugin",
      "version": "1.2.0",  // ← Must match plugin's plugin.json
      "description": "Updated description if changed"
      // ... other fields
    }
  ]
}

Critical: Keep Versions in Sync

• The version in marketplace.json MUST match the plugin's plugin.json version
• Inconsistencies break installation and updates

Why This Matters

• Discovery: Users browse marketplace.json to find plugins
• Installation: CLI uses marketplace.json to locate and install plugins
• Updates: Version tracking relies on marketplace.json
• Documentation: Plugin listings are generated from marketplace.json

Validation Scripts

validate-plugin.py

Location: {baseDir}/scripts/validate-plugin.py

Usage:

python3 {baseDir}/scripts/validate-plugin.py /path/to/plugin/

Validates:

1. Directory Structure

   • .claude-plugin/plugin.json exists
   • Referenced directories exist
   • README.md exists

2. plugin.json Schema

   • Valid JSON syntax
   • Required fields present
   • Name follows conventions
   • Version is semantic
   • Paths reference existing files

3. Component Validation

   • Agents: Valid YAML frontmatter and schema
   • Skills: Valid SKILL.md and directory structure
   • Commands: Valid YAML frontmatter
   • Hooks: Valid hooks.json schema

4. Security Checks

   • No exposed secrets in files
   • Safe script permissions
   • No dangerous bash operations

5. Documentation

   • README.md completeness
   • LICENSE file if license specified

Exit Codes:

• 0: All validations passed
• 1: Critical errors found
• 2: Warnings only (non-blocking)

Example Output:

✅ PLUGIN VALIDATION: my-plugin
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

📋 plugin.json
   ✓ Valid JSON syntax
   ✓ Required fields present
   ✓ Name follows conventions
   ✓ Semantic versioning

📁 Directory Structure
   ✓ .claude-plugin/plugin.json exists
   ✓ All referenced paths exist
   ✓ README.md exists

🔧 Components (5 total)
   ✓ 2 agents validated
   ✓ 1 skill validated
   ✓ 2 commands validated

🔒 Security
   ✓ No exposed secrets
   ✓ Safe script permissions

📝 Documentation
   ⚠ README.md missing usage examples

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
✅ VALIDATION PASSED (1 warning)

Security Considerations

When creating plugins:

1. Input Validation

   • Validate all command arguments
   • Sanitize file paths
   • Avoid command injection in scripts

2. Permissions

   • Use minimal allowed-tools in skills
   • Require user confirmation for destructive operations
   • Don't pre-approve dangerous tools (Bash, Write) unless necessary

3. Secrets Management

   • Never hardcode API keys, tokens, or credentials
   • Use environment variables: ${API_KEY}
   • Add .env to .gitignore
   • Document required environment variables

4. Script Safety

   • Validate script inputs
   • Avoid eval() and dynamic code execution
   • Use absolute paths, not relative
   • Set restrictive permissions (644 for files, 755 for executables)

5. Dependencies

   • Document all external dependencies
   • Pin versions for reproducibility
   • Avoid unnecessary dependencies

Best Practices

1. Start Simple, Iterate

Begin with minimal functionality:

• 1-2 components initially
• Test thoroughly
• Add features based on feedback
• Version bumps for each addition

2. Clear Documentation

Users should understand:

• What the plugin does (elevator pitch)
• How to install it
• How to use each component
• Configuration requirements
• Troubleshooting common issues

3. Consistent Naming

Use a naming scheme across components:

• plugin-name:category:action for namespaced commands
• Similar prefixes for related components
• Descriptive, not cute or clever

4. Version Thoughtfully

• Start at 1.0.0 for initial release
• Bump MAJOR for breaking changes
• Bump MINOR for new features
• Bump PATCH for bug fixes
• Update marketplace.json to match

5. Test Comprehensively

Before publishing:

• Manual testing of all components
• Edge case testing
• Integration testing
• Installation testing (clean environment)
• Validation script passing

6. Maintain Quality

After publishing:

• Monitor user feedback
• Fix bugs promptly
• Add features carefully
• Keep documentation updated
• Respond to issues

Reference Documentation

Comprehensive guides and examples:

• Plugin Architecture Guide

  • Design patterns and best practices
  • Component composition strategies
  • Scalability considerations

• Plugin Distribution Guide

  • Publishing to marketplace
  • Versioning strategies
  • Update workflows

• Plugin Examples

  • Real-world plugin examples
  • Common patterns and anti-patterns
  • Case studies

Your Role

When the user asks to create a plugin:

1. Assess Scope: Understand what the plugin should do
2. Recommend Architecture: Suggest component breakdown
3. Validate Approach: Ensure cohesive, not fragmented
4. Guide Creation: Use templates and validation
5. Ensure Quality: Comprehensive testing and documentation
6. Register Plugin: Update marketplace.json if applicable

Be proactive in:

• Recommending plugins over scattered components
• Suggesting cohesive component architectures
• Identifying security concerns early
• Ensuring comprehensive documentation
• Validating before considering "done"

Your goal is to help users create high-quality, well-structured plugins that provide real value and follow best practices.

Plugin Settings Pattern

Plugins can implement user-configurable settings using the .claude/plugin-name.local.md pattern. This allows users to customize plugin behavior on a per-project basis.

File Location and Format

Location: .claude/<plugin-name>.local.md in the project root

Format: YAML frontmatter + markdown body

---
# Plugin configuration (YAML frontmatter)
enabled: true
mode: strict
custom_option: value
---

# Plugin Context (markdown body)

Additional context or instructions that the plugin should consider.
This content can be loaded by hooks or skills.

Use Cases

1. Hook Activation Control

---
validation_enabled: true
auto_format: false
---

The hook script checks this setting:

#!/bin/bash
CONFIG_FILE=".claude/${PLUGIN_NAME}.local.md"

# Quick exit if config doesn't exist
[ ! -f "$CONFIG_FILE" ] && exit 0

# Parse enabled setting from frontmatter
ENABLED=$(sed -n '/^---$/,/^---$/p' "$CONFIG_FILE" | grep "^validation_enabled:" | cut -d: -f2 | tr -d ' ')

[ "$ENABLED" != "true" ] && exit 0

# Continue with hook logic...

2. Agent State Management

---
assigned_tasks:
  - review-api-endpoints
  - update-documentation
completed_reviews: 5
last_run: "2025-01-15"
---

3. Project-Specific Context

---
enabled: true
---

## Project Conventions

- Use TypeScript for all new code
- Follow the Airbnb style guide
- All API endpoints must have tests

## Domain Knowledge

This project manages customer billing. Key concepts:
- Subscriptions have monthly/annual cycles
- Invoices generate on billing dates

Parsing Settings in Hooks

Extract string/boolean fields:

get_setting() {
  local file="$1"
  local key="$2"
  sed -n '/^---$/,/^---$/p' "$file" | grep "^${key}:" | cut -d: -f2 | tr -d ' '
}

ENABLED=$(get_setting ".claude/my-plugin.local.md" "enabled")
MODE=$(get_setting ".claude/my-plugin.local.md" "mode")

Extract markdown body:

get_body() {
  local file="$1"
  sed '1,/^---$/d' "$file" | sed '1,/^---$/d'
}

CONTEXT=$(get_body ".claude/my-plugin.local.md")

Best Practices

1. Use .local.md suffix: Indicates user-local settings, should be in .gitignore
2. Provide sensible defaults: Plugin should work without settings file
3. Document all settings: List options in plugin README
4. Validate settings: Check for required values in hooks
5. Handle missing file: Always check if settings file exists first
6. Restart required: Settings only load at session start

Template: Plugin Settings File

---
# my-plugin settings
# Copy to .claude/my-plugin.local.md and customize

# Enable/disable the plugin for this project
enabled: true

# Validation strictness: strict | normal | lenient
mode: normal

# Custom options (plugin-specific)
option1: value1
option2: value2
---

# Project-Specific Context

Add any project-specific information here that the plugin should consider.

Integration with Components

In hooks (most common):

CONFIG=".claude/my-plugin.local.md"
[ -f "$CONFIG" ] && ENABLED=$(get_setting "$CONFIG" "enabled")

In skills (via description triggers): Skills can mention checking for project settings in their workflow.

In commands (via argument defaults): Commands can read settings for default values.

Common Questions

Q: When should I create a plugin vs individual components? A: Create a plugin when you have 3+ related components or want to distribute functionality as a package. Individual components are fine for one-off customizations.

Q: Can I include other plugins as dependencies? A: Not directly. Document required plugins in README.md and instruct users to install them separately.

Q: How do I handle plugin updates? A: Increment version in plugin.json, update marketplace.json, document changes in README.md, and test thoroughly before releasing.

Q: Can plugins have configuration files? A: Yes! Use .plugin-name.config.json or similar. Document configuration options in README.md.

Q: What's the difference between plugin keywords and categories? A: Keywords are for search (array of strings). Categories group plugins by type (single string). Both improve discoverability.

Q: How do I deprecate a plugin component? A: Document in README.md, add deprecation notice in component description, maintain for at least one MAJOR version, then remove and bump MAJOR version.

Summary

Creating plugins is about bundling expertise into reusable, distributable packages. Follow the structure, validate thoroughly, document comprehensively, and test extensively. Plugins should feel like natural extensions of Claude's capabilities, providing value without friction.