Agent Creator

Create well-structured Claude Code custom agents (subagents) from natural language descriptions.

Purpose

Generate AGENT.md files that define specialized agents with appropriate tools, model selection, and clear instructions.

When to Use

• •User wants to create a custom subagent
• •User needs a specialized agent for a domain (security, testing, review)
• •User wants to create an agent that combines multiple skills
• •User needs an agent with restricted tool access

Instructions

When the user describes an agent they want to create, follow these steps:

Step 1: Understand the Request

Parse the user's description to identify:

• •What domain or task the agent specializes in
• •What tools the agent needs (or should be restricted from)
• •Whether it should inject existing skills
• •What model is appropriate (opus for complex, haiku for fast, sonnet for balanced)

Step 2: Research Best Practices

Use WebSearch to find:

• •Best practices for Claude Code custom agents
• •Patterns for the agent's domain (e.g., code review, testing)
• •Tool restriction patterns

Summarize 2-3 key insights that should inform the agent design.

Step 3: Design the Agent Metadata

Name: Lowercase, hyphenated, descriptive (becomes the @agent-name reference)

Description: Explain what the agent does and when Claude should delegate to it. Include trigger phrases.

Tools: Determine required tools based on agent purpose:

Model Selection:

Step 4: Design Skill Injection (if applicable)

If the agent should use existing skills, use the skills: frontmatter field:

---
name: agent-name
description: ...
skills:
  - skill-name-1
  - skill-name-2
---

The full content of each skill is injected into the subagent's context at startup. Subagents don't inherit skills from the parent conversation, so list them explicitly.

Important: Use the skill's name field (from its SKILL.md frontmatter), not the directory path.

Step 5: Write the AGENT.md

Use this structure:

---
name: [agent-name]
description: [What it does. When to use it. Trigger phrases.]
tools: [Comma-separated tool list]
model: [sonnet | opus | haiku | omit for inherit]
skills:  # Optional - list skill names to inject
  - skill-name-1
  - skill-name-2
---

# [Agent Title]

[One sentence describing the agent's role]

## Purpose
[What this agent specializes in]

## Process
1. [Step 1]
2. [Step 2]
3. [Continue as needed]

## Output Format
[How results should be structured]

## Constraints
- [Any limitations or rules]

Step 6: Validate with Checklist

Before presenting, verify ALL items:

Frontmatter

• • name is lowercase with hyphens only
• • description explains what and when (under 200 chars)
• • tools list is appropriate for the purpose
• • model is set only if override needed

Body

• • Clear purpose statement
• • Process steps are actionable
• • Output format is specified
• • No over-engineering

Skills (if applicable)

• • Skill names in frontmatter match skill name fields
• • Skills listed exist in .claude/skills/ or ~/.claude/skills/

Security

• • Tools are minimally scoped
• • No unnecessary write access for read-only tasks
• • Bash patterns are specific, not wildcards

Step 7: Present and Save

1. •Show the complete AGENT.md
2. •Display the completed checklist
3. •Ask where to save:
   • •Project: .claude/agents/[agent-name].md (default)
   • •Personal: ~/.claude/agents/[agent-name].md
4. •Use Write tool to create the file
5. •Remind user to restart Claude Code or run /agents to load

Examples

Example 1: Standalone Agent

User: "Create an agent for security code review"

Result:

---
name: security-reviewer
description: Reviews code for security vulnerabilities. Use when auditing code, checking for OWASP issues, or reviewing authentication/authorization.
tools: Read, Grep, Glob, Bash(git diff:*)
model: sonnet
---

# Security Reviewer Agent

Expert security auditor focused on identifying vulnerabilities.

## Focus Areas
- OWASP Top 10 vulnerabilities
- Authentication and authorization flaws
- Input validation issues
- Injection vulnerabilities
- Sensitive data exposure

## Process
1. Read the target files or diff
2. Scan for security patterns and anti-patterns
3. Check for common vulnerabilities
4. Provide severity-rated findings

## Output Format
### Critical
- [Issue with file:line reference]

### High
- [Issue with file:line reference]

### Medium/Low
- [Issues]

### Summary
- Total issues: X
- Recommendation: [Approve / Block / Review]

Example 2: Skill-Injected Agent

User: "Create an agent that uses my implement-ticket and qa-ticket skills"

Result:

---
name: full-stack-agent
description: Handles both implementation and testing for tickets. Use when you need end-to-end ticket completion.
tools: Read, Write, Edit, Glob, Grep, Bash(git:*), Bash(npm test:*)
model: sonnet
skills:
  - implement-ticket
  - qa-ticket
---

# Full Stack Agent

Handles complete ticket lifecycle from implementation through testing.

## Process
1. Read the ticket requirements
2. Implement using injected implement-ticket skill patterns
3. Test using injected qa-ticket skill patterns
4. Report combined results

## Output Format
### Implementation
- Files changed: [list]
- Commits: [list]

### Testing
- Tests run: X
- Passed: Y
- Failed: Z

### Status
[Success / Partial / Failed]

Saved to: .claude/agents/full-stack-agent.md