Plugin Creator

Claude Code Plugin Creator

This skill helps you create well-structured Claude Code plugins following best practices from the official documentation and successful plugins like gemini-search.

When to Use This Skill

Invoke this skill when the user:

• Asks to create a new Claude Code plugin
• Wants to understand plugin structure and architecture
• Needs help with plugin.json, marketplace.json, or other config files
• Wants to add commands, agents, or hooks to their plugin
• Asks about plugin best practices or conventions

Plugin Architecture Overview

A Claude Code plugin consists of:

1. Core Configuration (.claude-plugin/)

   • plugin.json - Plugin metadata and entry points
   • marketplace.json - Marketplace listing information

2. Commands (commands/)

   • Markdown files defining slash commands (e.g., /search, /stats)
   • YAML frontmatter with name, description, usage, examples
   • Command execution instructions

3. Agents (agents/)

   • Markdown files defining subagents with isolated context
   • YAML frontmatter with description and capabilities
   • Agent-specific instructions and patterns

4. Hooks (hooks/)

   • hooks.json - Hook configuration with matchers
   • Bash scripts for PreToolUse, PostToolUse, UserPromptSubmit events
   • Event-driven automation

5. Scripts (scripts/)

   • Bash scripts for plugin functionality
   • Use ${CLAUDE_PLUGIN_ROOT} for portability

6. Documentation

   • README.md - User-facing docs
   • CLAUDE.md - Instructions for Claude Code
   • CHANGES.md - Changelog
   • CONTRIBUTING.md, TESTING.md, etc.

Step-by-Step Plugin Creation

Step 1: Initialize Plugin Structure

# Create directory structure
mkdir -p my-plugin/.claude-plugin
mkdir -p my-plugin/commands
mkdir -p my-plugin/agents
mkdir -p my-plugin/hooks
mkdir -p my-plugin/scripts
mkdir -p my-plugin/tests

# Initialize git repository
cd my-plugin
git init
git checkout -b main

Step 2: Create plugin.json

Create .claude-plugin/plugin.json:

{
  "name": "my-plugin",
  "version": "0.1.0",
  "description": "Brief description of what your plugin does",
  "author": "Your Name",
  "license": "MIT",
  "commands": "commands",
  "agents": "agents",
  "hooks": "hooks/hooks.json"
}

Important fields:

• name: kebab-case plugin identifier
• version: Semantic versioning (MAJOR.MINOR.PATCH)
• commands, agents, hooks: Relative paths to entry points

Step 3: Create marketplace.json

Create .claude-plugin/marketplace.json:

{
  "name": "my-plugin-marketplace",
  "owner": {
    "name": "Your Name",
    "email": "your.email@example.com"
  },
  "plugins": [
    {
      "name": "my-plugin",
      "source": {
        "type": "github",
        "repo": "username/repo-name"
      },
      "description": "Detailed description for the marketplace",
      "version": "0.1.0",
      "author": {
        "name": "Your Name"
      },
      "license": "MIT",
      "homepage": "https://github.com/username/repo-name",
      "repository": {
        "type": "git",
        "url": "https://github.com/username/repo-name.git"
      },
      "keywords": [
        "keyword1",
        "keyword2",
        "keyword3"
      ],
      "category": "productivity",
      "commands": "./commands",
      "agents": "./agents",
      "hooks": "./hooks/hooks.json",
      "strict": true,
      "features": [
        "Feature 1 description",
        "Feature 2 description",
        "Feature 3 description"
      ],
      "requirements": {
        "dependencies": [
          {
            "name": "dependency-name",
            "description": "What this dependency does",
            "install": "npm install -g dependency-name",
            "required": true
          }
        ]
      }
    }
  ]
}

Key sections:

• source.type: "github" or "directory" for local testing
• category: "productivity", "development", "utilities", etc.
• strict: Set to true for production plugins
• features: User-facing feature list
• requirements.dependencies: External tools needed

Step 4: Create Commands

Create a command file in commands/my-command.md:

---
name: my-command
description: Brief description of what this command does
usage: /my-command [arguments]
examples:
  - /my-command example argument
  - /my-command "quoted argument"
---

You are the command handler for the my-command in the my-plugin plugin.

## Execution Instructions

When this command is invoked:

1. Parse the user's arguments
2. Execute the required logic (call scripts, process data, etc.)
3. Present results to the user

## Example: Calling a Script

Run the following command:
```bash
bash "${CLAUDE_PLUGIN_ROOT}/scripts/my-script.sh" arg1 "{{USER_ARG}}"

Where {{USER_ARG}} is the argument provided by the user.

Response Format

Present results with:

• Clear summary of what was done
• Formatted output (tables, lists, etc.)
• Any warnings or errors
• Next steps or suggestions

Error Handling

If the command fails:

• Display clear error messages
• Suggest troubleshooting steps
• Recommend related commands

**Command best practices:**
- Use YAML frontmatter with required fields
- Provide clear usage examples
- Use `${CLAUDE_PLUGIN_ROOT}` for script paths
- Define clear error handling

### Step 5: Create Agents (Optional)

Create an agent file in `agents/my-agent.md`:

```markdown
---
description: Agent description and purpose
capabilities:
  [
    "capability-1",
    "capability-2",
    "capability-3",
  ]
---

# My Agent Name

This agent provides [specific functionality] with [key features].

## Features

- Feature 1 with context isolation
- Feature 2 with specific behavior
- Feature 3 with error handling

## Usage Patterns

Describe when and how this agent should be used:
- Pattern 1: [scenario]
- Pattern 2: [scenario]
- Pattern 3: [scenario]

## Important Notes

- Key architectural decision
- Performance considerations
- Limitations or constraints

Agent best practices:

• Use for context isolation and token savings
• Define clear capabilities
• Specify usage patterns
• Document architecture decisions

Step 6: Create Hooks (Optional)

Create hooks/hooks.json:

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "pattern|Pattern",
        "hooks": [
          {
            "type": "command",
            "command": "${CLAUDE_PLUGIN_ROOT}/hooks/pre-hook.sh"
          }
        ]
      }
    ],
    "PostToolUse": [
      {
        "matcher": "pattern|Pattern",
        "hooks": [
          {
            "type": "command",
            "command": "${CLAUDE_PLUGIN_ROOT}/hooks/post-hook.sh"
          }
        ]
      }
    ],
    "UserPromptSubmit": [
      {
        "matcher": "/command1|/command2",
        "hooks": [
          {
            "type": "command",
            "command": "${CLAUDE_PLUGIN_ROOT}/hooks/user-prompt-hook.sh"
          }
        ]
      }
    ]
  }
}

Hook types:

• PreToolUse: Before tool execution (suggestions, validation)
• PostToolUse: After tool execution (error detection, cleanup)
• UserPromptSubmit: On user command submission

Create hook script in hooks/pre-hook.sh:

#!/bin/bash

# Example hook script
# Receives JSON input via stdin with context about the event

# Read input
input=$(cat)

# Process input and provide feedback
echo "Hook executed successfully"

Hook best practices:

• Use matchers for targeted triggering
• Keep hooks fast and lightweight
• Use ${CLAUDE_PLUGIN_ROOT} for portability
• Handle errors gracefully

Step 7: Create Scripts

Create executable scripts in scripts/:

#!/bin/bash

# Script template
set -euo pipefail

# Use environment variables for configuration
SCRIPT_DIR="${CLAUDE_PLUGIN_ROOT}/scripts"
LOG_FILE="${LOG_FILE:-/tmp/my-plugin.log}"

# Function definitions
function main() {
    local arg1="$1"
    local arg2="${2:-default}"

    # Your logic here
    echo "Processing: $arg1"

    # Return JSON for structured output
    echo '{"status":"success","result":"data"}'
}

# Entry point
main "$@"

Script best practices:

• Use set -euo pipefail for safety
• Use ${CLAUDE_PLUGIN_ROOT} for paths
• Provide structured output (JSON)
• Log to /tmp/ for debugging
• Make scripts executable: chmod +x scripts/*.sh

Step 8: Create CLAUDE.md

Create CLAUDE.md for project-specific instructions:

# CLAUDE.md

This file provides guidance to Claude Code when working with this plugin.

## Project Overview

[Brief description of what this plugin does and its architecture]

## Common Commands

```bash
# Development & Testing
bash tests/run-tests.sh

# Using the plugin
/my-command [args]

Architecture & Key Concepts

Key Feature 1

[Explain important architectural decision]

Key Feature 2

[Explain another important concept]

Important Constraints

Do NOT:

• Thing to avoid 1
• Thing to avoid 2

Always:

• Best practice 1
• Best practice 2

### Step 9: Create Documentation

Essential docs:

1. **README.md** - User-facing documentation
   - Installation instructions
   - Usage examples
   - Features overview
   - Requirements

2. **CHANGES.md** - Changelog following Keep a Changelog format

3. **CONTRIBUTING.md** - Development workflow

4. **TESTING.md** - Testing guide

5. **LICENSE** - Open source license (MIT recommended)

### Step 10: Validate and Test

```bash
# Validate JSON files
jq empty .claude-plugin/plugin.json
jq empty .claude-plugin/marketplace.json
jq empty hooks/hooks.json

# Lint scripts
shellcheck scripts/*.sh hooks/*.sh

# Make scripts executable
chmod +x scripts/*.sh hooks/*.sh

# Test locally
# Add to .claude/settings.json:
{
  "extraKnownMarketplaces": {
    "local-test": {
      "source": {
        "source": "directory",
        "path": "/absolute/path/to/my-plugin"
      }
    }
  }
}

# Then in Claude Code:
/plugin add my-plugin@local-test

Best Practices

1. Context Isolation with Agents

Use agents for expensive operations to save tokens:

• Search operations
• API calls
• Heavy processing

Benefits: 30-40% token savings through context isolation

2. Caching Strategy

Implement caching for expensive operations:

CACHE_DIR="${CACHE_DIR:-/tmp/my-plugin-cache}"
CACHE_TTL="${CACHE_TTL:-3600}"  # 1 hour

cache_key=$(echo -n "$query" | md5sum | cut -d' ' -f1)
cache_file="$CACHE_DIR/$cache_key.json"

if [[ -f "$cache_file" ]]; then
    # Check if cache is fresh
    cache_age=$(($(date +%s) - $(stat -c %Y "$cache_file" 2>/dev/null || stat -f %m "$cache_file")))
    if [[ $cache_age -lt $CACHE_TTL ]]; then
        cat "$cache_file"
        exit 0
    fi
fi

3. Error Handling

Implement robust error handling:

set -euo pipefail

function error_handler() {
    local line_num=$1
    echo "Error on line $line_num" >&2
    exit 1
}

trap 'error_handler ${LINENO}' ERR

4. Logging

Log important events:

LOG_FILE="${LOG_FILE:-/tmp/my-plugin.log}"
ERROR_LOG="${ERROR_LOG:-/tmp/my-plugin-errors.log}"

function log() {
    echo "[$(date +'%Y-%m-%d %H:%M:%S')] $*" >> "$LOG_FILE"
}

function error_log() {
    echo "[$(date +'%Y-%m-%d %H:%M:%S')] ERROR: $*" >> "$ERROR_LOG"
}

5. Semantic Versioning

Follow semantic versioning strictly:

• MAJOR (X.0.0): Breaking changes
• MINOR (0.X.0): New features, backward compatible
• PATCH (0.0.X): Bug fixes, backward compatible

6. Security Considerations

• Never commit API keys or secrets
• Validate all user inputs
• Use secure temp file creation
• Follow principle of least privilege
• Log security-relevant events

7. Performance Optimization

• Use caching for repeated operations
• Implement context isolation with agents
• Stream large outputs
• Use efficient data structures
• Benchmark critical paths

8. Testing Strategy

Create tests/run-tests.sh:

#!/bin/bash
set -euo pipefail

echo "Running integration tests..."

# Test 1: JSON validation
jq empty .claude-plugin/plugin.json || exit 1
jq empty .claude-plugin/marketplace.json || exit 1

# Test 2: Script syntax
shellcheck scripts/*.sh || exit 1

# Test 3: Functional tests
bash scripts/my-script.sh test-arg || exit 1

echo "All tests passed!"

Common Patterns

Pattern 1: Command with Script Execution

Command file calls bash script, script returns JSON, command formats output.

Pattern 2: Agent-Based Processing

Main command spawns agent, agent processes in isolated context, results return to main.

Pattern 3: Hook-Driven Automation

Hooks detect patterns, trigger scripts, provide suggestions or automation.

Pattern 4: Caching Layer

Check cache first, execute on miss, store result, track analytics.

Marketplace Publishing

Prerequisites

1. GitHub repository with plugin code
2. Valid plugin.json and marketplace.json
3. Comprehensive README.md
4. LICENSE file (MIT recommended)
5. All tests passing

Publishing Steps

1. Prepare Release

   # Update version in all files
   # Update CHANGES.md
   git add -A
   git commit -m "chore: prepare release v0.1.0"
   git tag -a v0.1.0 -m "Release v0.1.0"
   git push origin main --tags
   

2. Create GitHub Release

   • Use GitHub UI or gh release create
   • Attach assets if needed
   • Copy changelog to release notes

3. Test Installation

   # Add marketplace to settings
   # Install plugin
   # Verify functionality
   

Troubleshooting

Plugin Not Loading

• Check JSON syntax with jq empty
• Verify paths in plugin.json
• Check file permissions
• Review Claude Code logs

Commands Not Working

• Verify YAML frontmatter format
• Check script paths and permissions
• Test scripts independently
• Review command error logs

Hooks Not Triggering

• Verify matchers are correct
• Check hooks.json syntax
• Test hook scripts independently
• Review hook execution logs

Resources

• Official Docs: https://docs.claude.com/en/docs/claude-code
• Skills Docs: https://docs.claude.com/en/docs/claude-code/skills
• Example Plugins: https://github.com/anthropics/skills
• Community: https://github.com/hesreallyhim/awesome-claude-code

Related Skills

This plugin-creator skill works together with other skills for a complete development workflow:

web-research Skill

Use the web-research skill to efficiently research plugin development topics:

• Latest Claude Code plugin best practices
• Example plugins and their implementations
• Marketplace requirements and guidelines
• Tool comparisons for plugin dependencies

The web-research skill uses the gemini-search agent for token-efficient searches with caching.

github-repo-management Skill

Use the github-repo-management skill for repository setup and CI/CD:

• Setting up GitHub Actions workflows for plugins
• Configuring issue templates and labels
• Release automation with semantic versioning
• Pull request workflows and branch protection
• GitHub Packages publishing

shell-script-quality Skill

Use the shell-script-quality skill for script quality assurance:

• ShellCheck linting for plugin scripts
• BATS testing framework setup
• CI/CD integration for shell script testing
• Pre-commit hooks for quality checks

Next Steps After Creating Plugin

1. Initialize git repository

   • Use the github-repo-management skill for repository setup
   • Configure branch protection and issue templates

2. Create comprehensive tests

   • Use the shell-script-quality skill for test setup
   • Add BATS tests for all scripts

3. Add CI/CD with GitHub Actions

   • Use the github-repo-management skill for workflow setup
   • Configure automated testing, linting, and releases

4. Research best practices

   • Use the web-research skill to find latest best practices
   • Research similar successful plugins

5. Write user documentation

   • README.md with installation and usage
   • CONTRIBUTING.md for contributors
   • CHANGES.md for changelog

6. Test locally before publishing

   • Local marketplace testing
   • Manual testing of all commands and hooks

7. Publish to marketplace

   • Create GitHub release
   • Verify installation from marketplace

8. Gather user feedback

   • Set up GitHub Discussions
   • Monitor issues and PRs

9. Iterate and improve

   • Address user feedback
   • Add new features based on usage patterns
   • Keep dependencies updated

Example: Complete Minimal Plugin

Here's a complete minimal plugin structure:

my-plugin/
├── .claude-plugin/
│   ├── plugin.json
│   └── marketplace.json
├── commands/
│   └── hello.md
├── scripts/
│   └── hello.sh
├── CLAUDE.md
├── README.md
├── CHANGES.md
└── LICENSE

With commands/hello.md:

---
name: hello
description: Say hello to the user
usage: /hello [name]
examples:
  - /hello World
---

Run: `bash "${CLAUDE_PLUGIN_ROOT}/scripts/hello.sh" "{{NAME}}"`
Present the result to the user.

And scripts/hello.sh:

#!/bin/bash
name="${1:-World}"
echo "Hello, $name!"

This is the simplest possible plugin. From here, add complexity as needed!