tomevault-io Create a new subagent by generating a Markdown file in .claude/agents/.
Process
1. Gather Requirements
If the user provided $ARGUMENTS, use it as the agent name. Otherwise ask:
- Name: lowercase, hyphens only (e.g.,
code-reviewer,test-runner) - Purpose: What task does this agent handle?
- Language: What language should the agent respond in? (default: Korean with English technical terms)
2. Determine Configuration
Based on the purpose, decide:
| Field | How to decide |
|---|---|
model |
haiku for fast search/lookup, sonnet for general tasks, opus for complex reasoning |
tools |
Only grant what's needed. Read-only tasks → Read, Glob, Grep, WebSearch, WebFetch. Code changes → add Edit, Write, Bash |
3. Write the Agent File
Create .claude/agents/{name}.md with this structure:
---
name: {name}
description: "{1-2 sentence description of when Claude should delegate to this agent. Include example triggers.}"
model: {model}
---
{System prompt: role definition, methodology, output format, guidelines}
4. Description Field Guidelines
The description field is critical — Claude uses it to decide when to auto-delegate. Write it to include:
- When to use this agent (trigger conditions)
- Example user messages that should activate it (use
<example>blocks) - Format examples in the pattern used by existing agents in
.claude/agents/
5. System Prompt Guidelines
The body after frontmatter is the agent's system prompt. Include:
- Role definition: Who is this agent? What expertise does it have?
- Core mission: One clear sentence on what it does
- Methodology: Step-by-step process the agent should follow
- Output format: How to structure the response
- Operational guidelines: Rules and constraints
- Language: Match the project convention (Korean with English technical terms)
6. Verify
After creating the file, read it back and confirm with the user.
Reference
Check existing agents in .claude/agents/ for style consistency before writing.
Converted and distributed by TomeVault — claim your Tome and manage your conversions.