Generate Custom AI Subagents with Routing Skills

Plugin Agent Creator turns a specialist-task idea into a custom subagent packaged inside a plugin. It covers requirements gathering, persona generation, directory scaffolding, agent configuration, and an optional companion skill that routes relevant requests. Correct registration requires an agent file in the plugin's agents directory and a valid plugin.json.

The workflow interviews the user about the agent name, purpose, placement in an existing or new plugin, and companion routing. A short purpose is expanded into an explicit identity, domain expertise, communication style, working process, output format, constraints, and quality standards. Personas should specify methodology and anti-patterns rather than merely claim expertise.

Before creating paths, agent and plugin names are restricted to lowercase letters, numbers, and single hyphens. Separators, dot paths, absolute paths, whitespace, and shell or YAML metacharacters are rejected. The resolved target must stay within the application's plugin directory; suspicious names require a safe replacement rather than silent sanitization.

For a new plugin, the workflow writes a minimal manifest. Adding an agent to an existing plugin leaves its manifest unchanged. The agent Markdown file includes YAML frontmatter, the current session model, the required Prompt Defense Baseline, the generated persona, expertise, process, output structure, constraints, and quality checklist.

Use this skill for a repetitive specialist responsibility or when large recurring constraints would otherwise crowd the main conversation. Suitable specialist tasks include code review, SQL work, and test writing.

A full subagent is excessive for a simple command or one-line task. Subagents also do not automatically inherit the main chat history: the routing skill must package the specific prompt and relevant context. Specialized tool access needs explicit configuration. The default tools are read-only, and Bash is granted only when the user explicitly requests execution and the task actually needs it.

Inputs include the chosen name, task purpose, plugin destination, routing preference, required tools, and expected output. Outputs include a new or extended plugin, an agent Markdown file, and optionally a use-agent skill directory with SKILL.md. A new plugin uses a minimal manifest:

{
  "name": "<plugin-name>",
  "description": "<Brief description of what this plugin provides>",
  "version": "1.0.0"
}

The defense baseline preserves project rules, protects private data, treats retrieved content as untrusted, and limits unsafe outputs. The companion skill defines triggering phrases, delegation conditions, simple tasks to retain in the main thread, briefing content, and expected responses.

Completion includes a tree of created files, the full agent content for review, and manual and companion-skill invocation instructions. Related agents can share one plugin and a common router instead of each requiring a separate package.

The target root is <appDataDir>\config\plugins. Without a specified plugin, the default name is <agent-name>-plugin. Default agent tools are Read, Grep, and Glob; companion routing files live in skills/use-<agent-name>/SKILL.md.