create-agent

Create a New Claude Code Subagent

You are creating a new custom subagent based on the user's request: $ARGUMENTS

Follow the steps below precisely to produce a production-quality subagent definition.

Step 1: Parse the Request

Extract from $ARGUMENTS:

• Agent name: The first token (use kebab-case, lowercase letters and hyphens only). If the user did not provide one, derive a concise name from the description.
• Purpose: Everything after the name describes what the agent should do.

Step 2: Determine Scope and Placement

Scope	Path	When to choose
Project	.claude/agents/<agent-name>.md	Agent is specific to this codebase and should be shared via version control.
Personal	~/.claude/agents/<agent-name>.md	Agent is useful across all projects but not team-shared.

Default to project scope (.claude/agents/<agent-name>.md) unless the user explicitly says "personal" or "global".

Step 3: Design the Agent Configuration

3a: Choose the Model

Model	Best for
haiku	Fast, read-only tasks like code search, exploration, quick lookups. Low cost.
sonnet	Balanced capability and speed. Good for analysis, reviews, moderate complexity.
opus	Maximum capability. Complex reasoning, architecture decisions, nuanced analysis.
inherit	Uses the same model as the main conversation. Default if omitted.

3b: Select Tools

Choose from Claude Code's available tools. Only grant what the agent needs:

Read-only tools (safe for exploration):

• Read - Read file contents
• Grep - Search file contents
• Glob - Find files by pattern

Modification tools (grant with care):

• Write - Write new files
• Edit - Modify existing files
• Bash - Execute shell commands

Special tools:

• Agent(type) - Spawn subagents of a specific type
• MCP server tools (reference configured server names)

3c: Configure Permission Mode

Mode	Behavior	Use when
default	Standard permission prompts	Interactive tasks needing oversight
acceptEdits	Auto-accept file edits	Trusted code modification tasks
dontAsk	Auto-deny permission prompts (allowed tools still work)	Background tasks that should not interrupt
bypassPermissions	Skip all permission checks	Fully trusted automation (use with extreme caution)
plan	Read-only exploration mode	Research and planning only

3d: Consider Advanced Options

Persistent Memory (memory field):

• user - Remember learnings across all projects
• project - Project-specific knowledge, shareable via version control
• local - Project-specific, not version controlled

Preloaded Skills (skills field):

• List skill names to inject their full content into the agent's context at startup

Max Turns (maxTurns), Background (background: true), Isolation (isolation: worktree), Hooks (hooks for PreToolUse/PostToolUse/Stop)

Step 4: Compose the Frontmatter

---
name: <agent-name>
description: <when Claude should delegate to this agent>
tools: <appropriate tool list>
model: <haiku|sonnet|opus|inherit>
permissionMode: <appropriate mode>
maxTurns: <optional limit>
skills: <optional list of skills to preload>
memory: <optional: user|project|local>
---

Step 5: Write the System Prompt

The markdown body after frontmatter becomes the agent's system prompt.

You are a [role description].

When invoked:
1. [Gather context]
2. [Analyze/process]
3. [Take action]
4. [Verify results]

[Domain guidelines, output format, constraints]

Step 6: Write the File

1. Create the agents directory if needed: mkdir -p .claude/agents
2. Write the agent file to <scope-path>/<agent-name>.md

Step 7: Validate and Report

Verify kebab-case naming, valid frontmatter, scoped tools, clear system prompt. Then tell the user the file path, how to invoke it, and example usage.