agent-development_skill

Overview

This skill guides creating, structuring, and validating autonomous agents for Claude Code plugins. It explains file format, frontmatter fields, system prompt design, triggering examples, and testing workflows. Use it to produce reliable, least-privilege agents that trigger only for intended situations.

How this skill works

The skill inspects the agent markdown layout: required YAML frontmatter (name, description, model, color, optional tools) and the body that becomes the system prompt. It enforces identifier rules, description with concrete <example> blocks, model/color choices, and tool restrictions, then runs validation and testing steps to confirm triggering and output formats. It also provides templates and an AI-assisted prompt pattern to accelerate agent creation.

When to use it

• When you need to create a new autonomous agent for a Claude Code plugin
• When you want to add concrete triggering examples and frontmatter to an existing agent
• When you must design a system prompt with responsibilities, process steps, and output format
• When you need to limit tool access or pick a distinct color for UI clarity
• When you want to validate identifiers, descriptions, and test agent triggers

Best practices

• Write the description in the required 'Use this agent when...' format and include 2–4 <example> blocks showing context, user input, assistant response, and commentary
• Choose an identifier with lowercase letters and hyphens (3–50 chars) and follow validation rules strictly
• Use model: inherit unless a specific capability justifies sonnet/opus/haiku
• Limit tools to the minimum required to follow least-privilege principles
• Write the system prompt in second person with clear responsibilities, step-by-step analysis, quality standards, edge-case handling, and a precise output format
• Test triggering with varied phrasings and validate outputs against the defined format

Example use cases

• Create a code-reviewer agent that runs read-only analysis and returns a structured checklist
• Add a test-generator agent with tools ["Read","Bash","Grep"] and examples showing user prompts that should trigger it
• Write a security-analyzer agent that uses color red and restricts tools for sensitive checks
• Generate an agent-creation helper using the AI-assisted prompt pattern to produce identifier, description, and system prompt JSON