plugin-creator

Plugin Creator

This skill is the entry point for creating Claude Code plugins. It describes the canonical plugin layout and points to the focused authoring skills used by the plugin-creator agent.

Canonical plugin layout

<plugin-name>/
├── .claude-plugin/plugin.json   # required manifest
├── agents/<name>.md             # optional subagents
├── commands/<name>.md           # optional slash commands
├── hooks/hooks.json             # optional hook configuration
└── skills/<name>/SKILL.md       # optional skills

Manifest essentials (.claude-plugin/plugin.json)

• name (required, kebab-case, matches the plugin directory).
• version, description, author, repository (strongly recommended).
• Component arrays — include only those the plugin ships:
  • agents: list of .md paths.
  • skills: list of skill directory paths.
  • commands: list of .md paths.
  • hooks: path to a hooks.json file.

Every path is relative to the plugin root and every referenced file must exist.

Authoring workflow

1. Decide the plugin name and which components are needed.
2. Create the directories and the manifest.
3. Author each component using the dedicated skill:
   • Agents → writing-agents
   • Hooks → writing-hooks
   • Skills → writing-skills
   • Commands → writing-commands
4. Validate that plugin.json parses and every referenced file exists.
5. Tell the user the install commands:

   /plugin marketplace add <owner>/<repo>
   /plugin install <plugin-name>
   

When to use which component

• Agent — a focused subagent with its own system prompt and tool scope.
• Skill — reusable, model-invoked knowledge/instructions Claude loads when relevant.
• Command — a user-invoked slash command (/name).
• Hook — a deterministic shell command triggered by a Claude Code lifecycle event (PreToolUse, PostToolUse, UserPromptSubmit, SessionStart, Stop, Notification, ...).