name: create description: Create plugin components (commands, skills, hooks, agents, prompts) interactively allowed-tools: [Read, Write, Edit, Glob, Grep, AskUserQuestion, Bash, Task]

Create Plugin Component

Scaffold plugin components through iterative interviews. Works with any Claude Code plugin.

Expert Consultation

Always consult the claude-code-guide agent when:

• Designing hook logic (exit codes, matchers, event types)
• Choosing between component types (skill vs agent vs hook)
• Validating generated content against best practices
• User asks questions about Claude Code features

task:
  subagent_type: claude-code-guide
  prompt: |
    User is creating a <component_type> named <name>.

    Validate this design against Claude Code best practices:
    - <design details>

    Flag any anti-patterns or suggest improvements.

Context

!ls .claude-plugin/plugin.json 2>/dev/null && echo "Plugin detected" || echo "No plugin detected - will create in current directory"

Plugin Root Detection

Detect the target plugin directory:

1. If $CLAUDE_PLUGIN_ROOT set, use it
2. If .claude-plugin/plugin.json exists, use current directory
3. Otherwise, ask user for target path or create new plugin structure

Phase 1: Component Type

Parse argument if provided (/create command, /create skill, etc.).

If no argument or argument not recognized:

question: "What type of component do you want to create?"
header: "Component"
options:
  - label: "Command"
    description: "Slash command entry point (/plugin:name)"
  - label: "Skill"
    description: "Reusable workflow or knowledge"
  - label: "Hook"
    description: "Event-driven automation script"
  - label: "Agent"
    description: "Autonomous subagent definition"
  - label: "Prompt"
    description: "MCP prompt template with arguments"
multiSelect: false

Phase 2: Type-Specific Interview

Based on component type, follow the detailed reference:

Phase 2.5: Entity Scan

Before finalizing, scan all existing Claude entities for conflicts and integration opportunities.

Checklist

• Naming conflicts: Check all entity locations for the chosen name
• Related components: Find similar functionality for cross-referencing
• Hook opportunities: Identify potential integrations
• Expert consultation: If conflicts or complex integrations found

Check for Naming Conflicts

Use Glob to check for existing entities with the same name:

1. Glob("commands/<name>.md") - Existing command
2. Glob("skills/<name>/SKILL.md") - Existing skill
3. Glob("hooks/<name>.sh") - Existing hook script
4. Glob("agents/<name>.md") - Existing agent
5. Glob("prompts/<name>.md") - Existing prompt

Also check hooks/hooks.json for hook registrations using the same name pattern.

Find Related Components

Use Grep to find related functionality by keywords from the new component's name/description:

1. Grep(pattern: "<keywords>", path: "commands/") - Related commands
2. Grep(pattern: "<keywords>", path: "skills/") - Related skills
3. Grep(pattern: "<keywords>", path: "hooks/") - Related hooks
4. Grep(pattern: "<keywords>", path: "agents/") - Related agents

Integration Analysis

Based on component type being created:

If creating command/skill:

• Check if existing hooks could enhance it (e.g., PostToolUse notifications)
• Identify commands that should reference this in their docs

If creating hook:

• Identify commands/skills it should integrate with
• Check for related hooks that might conflict

If conflicts or complex integrations found:

• Consult claude-code-guide for naming conventions and resolution

Output

Report findings to user before Phase 3 confirmation:

## Entity Scan Results

✓ No naming conflicts (or ⚠ Conflicts: <list>)

**Related Components** (if any):
- commands/foo.md - Similar functionality [REVIEW]
- hooks/bar.sh - Could trigger after this command [CONSIDER]

**Integration Suggestions**:
- Reference this from related-command.md
- Consider PostToolUse hook for notifications

Phase 3: Expert Review & Confirmation

For hooks: Before showing preview, consult claude-code-guide to validate:

• Exit code usage (0=allow, 2=block)
• Defensive stdin pattern
• stop_hook_active check for Stop hooks
• Proper use of ${CLAUDE_PLUGIN_ROOT}

Before writing files, show preview:

## Component Preview

**Type**: <type>
**Name**: <name>
**Files to create**:
- <file 1>
- <file 2>

### Content Preview

<show generated content>

Proceed with creation?

question: "Create this component?"
header: "Confirm"
options:
  - label: "Yes, create it"
    description: "Write files and complete setup"
  - label: "Edit first"
    description: "Make changes before creating"
  - label: "Cancel"
    description: "Don't create anything"
multiSelect: false

Phase 4: Write Files

1. Create necessary directories
2. Write component files
3. For hooks: update hooks.json
4. For command hooks: make script executable

Constraints

• Consult claude-code-guide: For hooks and complex components, spawn the expert agent to validate design
• Detect plugin root: Use $CLAUDE_PLUGIN_ROOT or .claude-plugin/plugin.json
• Safe file creation: Never overwrite without asking
• Valid names: Enforce kebab-case for all names
• Executable hooks: chmod +x for shell scripts
• hooks.json merge: Preserve existing hooks when adding new ones
• Best practices: Use defensive stdin pattern, proper exit codes, ${CLAUDE_PLUGIN_ROOT} paths