Create Agent

Create Agent Skill

This skill guides you through creating a well-structured Claude agent with proper frontmatter, system prompt, and examples.

When to Use This Skill

Activate this skill when the user wants to:

• Create a new Claude agent
• Customize Claude's behavior for specific tasks
• Structure an agent definition file
• Understand agent components (frontmatter, prompts, examples)

Agent Creation Workflow

1. Understand Agent Purpose

Ask the user (or infer from context):

• What is the agent's primary role/expertise?
• What specific tasks should it handle?
• Does it need to be read-only or can it write files?
• What level of capability is needed? (quick responses vs deep analysis)

2. Determine Agent Configuration

Model Selection:

• opus - Complex reasoning, deep analysis, architectural decisions
• sonnet - Balanced tasks, code review, general development (most common)
• haiku - Quick responses, simple tasks, high volume

Tool Selection (based on agent role):

• Read-only agents: Read, Grep, Glob
• Research agents: + mcp__context7__*, mcp__exa__*, WebSearch, WebFetch
• Code agents: + Write, Edit, mcp__ide__*, mcp__lsp-api__*
• Review agents: + mcp__ide__getDiagnostics, mcp__github__*
• Interactive agents: + AskUserQuestion, TodoWrite

Color Selection (for UI identification):

• cyan - Architecture/design
• purple - Code review/quality
• green - Testing/verification
• blue - Documentation/writing
• red - Debugging/troubleshooting
• yellow - Research/exploration

3. Generate Agent File

Create a markdown file with this structure:

---
name: descriptive-name
description: Brief 1-2 sentence description of when to use this agent
model: sonnet
color: purple
tools:
  - Read
  - Grep
  - Glob
  - [additional tools as needed]
---

# Role: [Agent Name]

[2-3 sentence description of agent's role and expertise]

## Core Responsibilities

1. **[Responsibility 1]** - [Description]
2. **[Responsibility 2]** - [Description]
3. **[Responsibility 3]** - [Description]

## [Methodology/Process Section]

[Detailed instructions on how the agent should approach tasks]

## Success Criteria

A successful [outcome] should:
- [ ] [Criterion 1]
- [ ] [Criterion 2]
- [ ] [Criterion 3]

4. Validation Checklist

Before saving, verify:

• YAML frontmatter is valid (use triple dashes ---)
• Name is lowercase with hyphens (e.g., code-reviewer)
• Description is concise and actionable (< 200 chars)
• Model is appropriate for task complexity
• Tools match agent capabilities
• System prompt defines clear role
• Core responsibilities are specific (not vague)
• Success criteria are measurable

Agent File Naming

Save agent files as:

• .claude/agents/[name].md (for local project)
• Or in your agent package's templates/claude/agents/ directory

Use descriptive names that match the name field in frontmatter.

Quick Start Templates

For common agent types, consider using these templates as starting points:

• templates/basic-agent.md - Generic structure
• templates/research-agent.md - Read-only with research tools
• templates/code-agent.md - Implementation with write access
• templates/review-agent.md - Code review and quality

Access templates with: "Use the research agent template"

Reference Materials

For detailed information, consult:

• reference.md - Complete agent structure documentation
• examples.md - Real-world agent examples with pattern analysis

Tips for Effective Agents

1. Be Specific: Clear, focused role beats general-purpose
2. Match Model to Task: Don't use opus for simple tasks (cost consideration)
3. Appropriate Tools: Only include tools agent actually needs
4. Actionable Prompts: System prompt should give clear behavioral guidance
5. Test with Examples: Try agent with realistic scenarios after creation
6. Iterate: Refine based on actual usage patterns

Common Patterns

Read-Only Analysis Agent:

• Model: sonnet
• Tools: Read, Grep, Glob, mcp__context7__*
• Use case: Code review, documentation analysis

Implementation Agent:

• Model: sonnet or opus
• Tools: Read, Write, Edit, Grep, Glob
• Use case: Feature development, refactoring

Research Agent:

• Model: sonnet
• Tools: Read, Grep, mcp__context7__*, mcp__exa__*, WebSearch
• Use case: Library research, best practices investigation

Interactive Planning Agent:

• Model: opus
• Tools: Read, Grep, Glob, AskUserQuestion, TodoWrite
• Use case: Architecture decisions, technical planning

Example Invocations

Here are examples of how users might request different agents:

• "Create a code review agent for TypeScript"
• "I need an agent that helps with system architecture"
• "Make a debugging agent for Go services"
• "Create a test-driven development agent"
• "Build an agent for API documentation"

For each request, determine the appropriate model, tools, and structure based on the patterns above.