agent-design_skill

Overview

This skill helps design AI agents for the CMAT multi-agent system by defining clear roles, selecting appropriate tools and skills, and drafting focused personas. It produces agent specifications that map tasks to capabilities and expected outputs to ensure reliable software development support. The goal is fast, repeatable creation of agents that do one thing well and collaborate effectively.

How this skill works

Start by clarifying the task and scope the agent must cover. Map required capabilities to concrete tools (file ops, search, edit, run, web fetch) and pick domain skills. Write a short persona with responsibilities, examples, and a defined output format so the agent can be instantiated or tested within Claude Code.

When to use it

• Creating a new agent for a CI/CD, code review, or testing workflow
• Defining responsibilities and tool access for an existing agent
• Choosing which tools and skills an agent needs for a software task
• Designing collaboration patterns and handoffs between multiple agents
• Refining agent prompts, outputs, or persona instructions after testing

Best practices

• Give each agent a single, clear responsibility to minimize scope overlap
• Match tools to actual needs (avoid granting every tool by default)
• Include domain skills relevant to the task (e.g., api-design, testing)
• Provide concrete examples, checklists, and an explicit output format
• Keep persona instructions focused and concise (avoid vague directives)

Example use cases

• Design an API Reviewer agent that reads specs, greps code, and writes a review
• Create a Test Runner agent with Bash and Edit tools to run and fix failing tests
• Build a Dependency Auditor that web-fetches advisories and reports vulnerabilities
• Define a Release Manager agent to assemble changelogs and run release scripts
• Compose a Multi-Agent pipeline where a Linter agent flags issues and a Remediator agent patches them