Agent Skills: Claude Plugin Development

Claude Plugin Development

Complete lifecycle for developing, validating, and distributing Claude Code plugins.

Steps

1. Define plugin scope and components needed
2. Initialize plugin structure with plugin.json
3. If adding commands, load the outfitter:claude-commands skill
4. If adding agents, load the outfitter:claude-agents skill
5. If adding hooks, load the outfitter:claude-hooks skill
6. If adding skills, load the outfitter:skills-dev skill
7. Delegate by loading the outfitter:claude-plugin-audit skill for validation
8. Fix issues and distribute

Quick Start

# 1. Scaffold plugin
./scripts/scaffold-plugin.sh my-plugin --with-commands

# 2. Add components (commands, agents, hooks, skills)
# 3. Test locally
/plugin marketplace add ./my-plugin
/plugin install my-plugin@my-plugin

# 4. Distribute
git push origin main --tags

Lifecycle Overview

Discovery -> Init -> Components -> Validate -> Distribute -> Marketplace
    |         |          |            |            |             |
    v         v          v            v            v             v
 Purpose   Scaffold   Commands    Structure    Package      Catalog
  Scope    plugin.json  Agents     Testing     Version      Publish
  Type      README      Hooks      Quality     Release       Share

Stage 1: Discovery

Before creating a plugin, clarify:

| Question | Impact | |----------|--------| | What problem does this solve? | Plugin scope and features | | Who will use it? | Distribution method | | What components are needed? | Commands, agents, hooks, MCP servers | | Where will it live? | Personal, project, or marketplace |

Stage 2: Initialization

Standalone Plugin

Standalone plugins need their own .claude-plugin/plugin.json:

my-plugin/
├── .claude-plugin/
│   └── plugin.json      # Required for standalone
├── README.md            # Required for distribution
├── commands/            # Optional components
├── agents/
├── skills/
└── hooks/

plugin.json (Standalone)

{
  "name": "my-plugin",
  "version": "1.0.0",
  "description": "Brief description of what this plugin does",
  "author": {
    "name": "Your Name",
    "email": "you@example.com"
  },
  "license": "MIT"
}

Marketplace with Local Plugins (Consolidated)

For marketplaces where all plugins live in the same repo, use strict: false to consolidate metadata. Plugins don't need their own manifests:

my-marketplace/
├── .claude-plugin/
│   └── marketplace.json # All metadata here (strict: false)
├── plugin-a/
│   └── commands/
├── plugin-b/
│   └── skills/
└── README.md

marketplace.json (Consolidated)

{
  "name": "my-marketplace",
  "owner": {
    "name": "Team Name",
    "email": "team@example.com"
  },
  "strict": false,
  "plugins": [
    {"name": "plugin-a", "source": "./plugin-a", "version": "1.0.0", "description": "Plugin A", "license": "MIT"},
    {"name": "plugin-b", "source": "./plugin-b", "version": "1.0.0", "description": "Plugin B", "license": "MIT"}
  ]
}

Benefits: Single source of truth, no version drift between marketplace and plugin manifests.

For external plugins (GitHub repos), use minimal entries and let the external repo own its manifest.

See structure.md for complete plugin.json schema.

Stage 3: Components

Add components based on plugin needs. See Steps section for which skills to load.

Slash Commands

Create custom commands in commands/ directory:

---
description: "Review code for quality issues"
---

Review the following code: {{0}}

Check for: code style, bugs, performance, security

For complex commands, load the outfitter:claude-commands skill.

Custom Agents

Define specialized agents in agents/ directory:

---
name: security-reviewer
description: "Security-focused code reviewer"
---

You are a security expert. When reviewing code:
1. Check for vulnerabilities
2. Verify input validation
3. Report issues with severity levels

For agent design patterns, load the outfitter:claude-agents skill.

Event Hooks

Two ways to define hooks:

File-based (auto-discovered from hooks/hooks.json):

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

Inline in plugin.json - same structure, add "hooks" key directly.

Hook types: PreToolUse, PostToolUse, UserPromptSubmit, Stop, SessionStart, SessionEnd

For hook implementation, load the outfitter:claude-hooks skill. See structure.md for hook JSON format and script interface.

Skills

Add reusable methodology patterns in skills/ directory. For skill authoring, load the outfitter:skills-dev skill.

MCP Servers

{
  "mcpServers": {
    "my-server": {
      "command": "${CLAUDE_PLUGIN_ROOT}/servers/my-server",
      "args": ["--config", "${CLAUDE_PLUGIN_ROOT}/config.json"],
      "env": {"API_KEY": "${MY_API_KEY}"}
    }
  }
}

Path variables: ${CLAUDE_PLUGIN_ROOT} (plugin directory), ${VAR_NAME} (env var)

Plugin Caching

When plugins are installed, Claude Code copies them to a cache directory. This has implications:

• Path traversal breaks: ../../shared/file.md will not work after install
• Keep resources inside plugin: Shared scripts, rules, and assets must be within plugin directory
• Cross-plugin dependencies: Use skill invocation (plugin:skill-name) instead of file references

See caching.md for workarounds and best practices.

Stage 4: Validation

Before distribution, validate the plugin.

Checklist

Structure:

• [ ] Standalone: plugin.json exists and is valid JSON
• [ ] Marketplace (consolidated): metadata in marketplace.json with strict: false
• [ ] Required fields present (name, version, description)
• [ ] Plugin name matches directory name (kebab-case)

Components:

• [ ] Commands have YAML frontmatter with description
• [ ] Agents have YAML frontmatter with name and description
• [ ] Hook scripts are executable (chmod +x)
• [ ] Hook matchers are valid regex

Documentation:

• [ ] README.md with installation instructions
• [ ] LICENSE file included

Local Testing

# Add as local marketplace
/plugin marketplace add ./my-plugin

# Install and test
/plugin install my-plugin@my-plugin

# Test commands
/my-command arg1 arg2

See structure.md for validation commands and detailed component schemas.

Stage 5: Distribution

Semantic Versioning

Follow semver (MAJOR.MINOR.PATCH):

• MAJOR: Breaking changes
• MINOR: New features (backward compatible)
• PATCH: Bug fixes

Release Workflow

# 1. Update version in plugin.json
# 2. Update CHANGELOG.md
# 3. Commit and tag
git add plugin.json CHANGELOG.md
git commit -m "chore: release v1.0.0"
git tag v1.0.0
git push origin main --tags

# 4. Create GitHub release
gh release create v1.0.0 --title "v1.0.0" --notes "Initial release"

Distribution Methods

| Method | Best For | Setup | |--------|----------|-------| | GitHub repo | Public/team plugins | Push to GitHub | | Git URL | GitLab, Bitbucket | Full URL in source | | Local path | Development/testing | Relative path |

See distribution.md for packaging, CI/CD, and release automation.

Stage 6: Marketplace

A marketplace catalogs plugins for discovery and installation.

Creating a Marketplace

Create .claude-plugin/marketplace.json:

{
  "name": "my-marketplace",
  "owner": {"name": "Team Name", "email": "team@example.com"},
  "plugins": [
    {"name": "my-plugin", "source": "./plugins/my-plugin"}
  ]
}

Plugin Sources

// Relative path
{"source": "./plugins/my-plugin"}

// GitHub
{"source": {"source": "github", "repo": "owner/plugin-repo", "ref": "v1.0.0"}}

// Git URL
{"source": {"source": "url", "url": "https://gitlab.com/team/plugin.git"}}

Commands

/plugin marketplace add owner/repo       # Add marketplace
/plugin marketplace list                  # List available
/plugin install plugin-name@marketplace  # Install from marketplace
/plugin marketplace update marketplace   # Update

See marketplace.md for full schema, team configuration, and hosting strategies.

Best Practices

Naming Conventions

• Plugin name: kebab-case (e.g., dev-tools)
• Commands: kebab-case (e.g., review-pr)
• Agents: kebab-case (e.g., security-reviewer)

Security

• Never hardcode secrets in plugin files
• Use environment variables for sensitive data
• Validate all user inputs in hooks
• Document security requirements

Documentation

• README.md: Overview, installation, usage examples
• CHANGELOG.md: Version history with semver
• LICENSE: Appropriate license file

Troubleshooting

Plugin not loading:

• Standalone: verify plugin.json syntax: jq empty .claude-plugin/plugin.json
• Marketplace: verify marketplace.json syntax and strict: false if no plugin.json
• Check plugin name matches directory
• Ensure required fields present (name, version, description)

Commands not appearing:

• Verify YAML frontmatter exists
• Check files in commands/ directory

Hooks not executing:

• Check scripts executable: chmod +x
• Verify matcher regex correct
• Test hook script independently

MCP servers failing:

• Verify server binary exists
• Check environment variables set
• Review logs: ~/Library/Logs/Claude/

• structure.md - Directory layout, plugin.json schema, component formats
• distribution.md - Packaging, versioning, CI/CD, release automation
• marketplace.md - Marketplace schema, hosting, team configuration
• caching.md - Plugin caching behavior and cross-plugin dependencies

ALWAYS:

• Standalone plugins: create .claude-plugin/plugin.json
• Marketplace local plugins: use strict: false and consolidate metadata in marketplace.json
• External plugins: let the external repo own its manifest
• Keep plugin resources within plugin directory (caching limitation)
• Use kebab-case for all names
• Include README.md and LICENSE for distribution
• Follow semantic versioning

NEVER:

• Hardcode secrets in plugin files
• Use path traversal (../) for cross-plugin resources
• Skip validation before distribution
• Omit description (in plugin.json or marketplace entry)

Related Skills

• claude-commands - Slash command development
• claude-agents - Custom agent design
• claude-hooks - Event hook implementation
• skills-dev - Skill creation patterns