🚨 CRITICAL GUIDELINES

Windows File Path Requirements

MANDATORY: Always Use Backslashes on Windows for File Paths

When using Edit or Write tools on Windows, you MUST use backslashes (\) in file paths, NOT forward slashes (/).

Examples:

• ❌ WRONG: D:/repos/project/file.tsx
• ✅ CORRECT: D:\repos\project\file.tsx

This applies to:

• Edit tool file_path parameter
• Write tool file_path parameter
• All file operations on Windows systems

Documentation Guidelines

NEVER create new documentation files unless explicitly requested by the user.

• Priority: Update existing README.md files rather than creating new documentation
• Repository cleanliness: Keep repository root clean - only README.md unless user requests otherwise
• Style: Documentation should be concise, direct, and professional - avoid AI-generated tone
• User preference: Only create additional .md files when user specifically asks for documentation

---

Create Plugin

Autonomously create production-ready Claude Code plugins with complete structure, documentation, and marketplace packaging.

Purpose

Guides Claude through the complete plugin creation workflow: fetching latest docs, generating all files (plugin.json, commands, agents, skills, README), creating marketplace structure, and providing installation instructions.

Instructions

1. Fetch latest documentation from docs.claude.com (plugins-reference, plugin-marketplaces)
2. Detect repository context - Check git config for author info and marketplace.json existence
3. Infer requirements from user request - Only ask questions if genuinely ambiguous
4. Create comprehensive structure:
   • Plugin manifest with all appropriate metadata fields
   • Commands, agents, and Agent Skills as needed (2025: Skills are auto-discovered from skills/ directory)
   • Hooks for automated workflows (PreToolUse, PostToolUse, SessionStart, etc.)
   • MCP servers for external integrations (inline in plugin.json or .mcp.json)
   • Complete README with examples and installation instructions
   • Marketplace structure if not in existing marketplace repo
5. Apply 2025 best practices:
   • Use Agent Skills for dynamic knowledge loading
   • Configure hooks for automated validation and testing
   • Support repository-level plugin configuration via .claude/settings.json
   • Use ${CLAUDE_PLUGIN_ROOT} environment variable for portable paths
6. Validate plugin.json schema - CRITICAL RULES:
   • ✅ repository must be a string URL, NOT an object
   • ✅ agents field is NOT needed - auto-discovered from agents/ directory
   • ✅ skills field is NOT needed - auto-discovered from skills/ directory
   • ✅ commands field is optional - auto-discovered from commands/ directory by default
   • ✅ Only include: name, version, description, author (object), homepage, repository (string), license, keywords (array)
   • ❌ NEVER include agents: {...} or skills: {...} in plugin.json - these cause validation errors
7. 🚨 MANDATORY: Register in marketplace.json (NON-NEGOTIABLE):
   • FIRST: Check if .claude-plugin/marketplace.json exists at repo root
   • IF EXISTS: Read it and add new plugin entry to plugins array
   • FORMAT: {"name": "plugin-name", "source": "./plugins/plugin-name", "description": "...", "version": "1.0.0", "author": {"name": "..."}, "keywords": [...]}
   • SYNCHRONIZE: Descriptions must match between plugin.json, marketplace.json, and README.md
   • ⚠️ PLUGIN CREATION IS NOT COMPLETE UNTIL REGISTERED IN MARKETPLACE.JSON ⚠️
8. Update README.md - Add plugin to appropriate category
9. Provide clear instructions for GitHub-first installation and repository-level configuration

Best Practices

Plugin Design Philosophy (2025)

Agent-First, Minimal Commands:

• Primary interface: Expert agent that users interact with conversationally
• Minimal commands: Only 0-2 slash commands per plugin for specific, high-value workflows
• Why: Users want to invoke domain expertise, not navigate command menus
• Commands only for: Automated workflows, batch operations, or specialized utilities
• Agent handles: Questions, guidance, code generation, troubleshooting, best practices

Agent Naming Standard:

• CRITICAL: Every plugin must have exactly ONE primary agent named {domain}-expert
• Pattern: docker-master → agent named docker-expert
• Pattern: terraform-master → agent named terraform-expert
• Why: Predictable names allow Claude to reliably invoke the correct agent
• Never: Use multiple specialized agents or non-standard names

Examples:

• ✅ dotnet-microservices-master: 0 commands, 1 agent (dotnet-microservices-expert)
• ✅ docker-master: 0 commands, 1 agent (docker-expert)
• ❌ OLD: 10+ commands + multiple agents - overwhelming and unpredictable!

Default to action, not questions - infer intent from context

• Include agent and Agent Skills by default
• Commands only when genuinely needed for automation
• Use detected git config values for author fields (never use placeholders)
• Create in plugins/ subdirectory if marketplace.json exists at repo root
• ALWAYS update .claude-plugin/marketplace.json when adding plugins to a marketplace repository
• Synchronize descriptions and keywords between plugin.json, marketplace.json, and README.md
• Add hooks for common automation needs (testing, linting, validation)
• Use ${CLAUDE_PLUGIN_ROOT} for all internal paths (scripts, configs, MCP servers)
• Include .claude/settings.json template for team distribution
• Leverage Agent Skills for dynamic, context-efficient knowledge delivery
• Configure MCP servers inline in plugin.json for simpler distribution
• Emphasize GitHub marketplace installation for cross-platform reliability (especially Windows/Git Bash)
• Document shell detection for plugin testing ($MSYSTEM for Git Bash/MinGW)
• Include path conversion guidance for Git Bash users developing plugins
• Support repository-level automatic installation for team standardization

Example Usage

/create-plugin for Git workflow automation
/create-plugin with deployment commands and rollback features
/create-plugin that helps with code reviews and security scanning

The plugin-master skill activates automatically to provide complete templates and current best practices.