creating-agents

# Create Agent

Guide the user through creating a highly effective Claude Code agent.

**Important:** Use "ultrathink" extended thinking for agent design decisions.

## Workflow

### Phase 1: Discovery

Ask the user clarifying questions to understand the agent requirements:

1. **Purpose:** What specialized role should this agent fill? (e.g., code reviewer, test runner, documentation writer)
2. **Scope:** Should it be project-specific (`.claude/agents/`) or personal (`~/.claude/agents/`)?
3. **Delegation:** Should Claude delegate to it proactively, or only when the user explicitly requests it?
4. **Tools:** What tools does this agent need? (Read, Write, Edit, Bash, Glob, Grep, WebSearch, WebFetch, etc.)
5. **Model:** What model should it use? (`opus` for complex reasoning, `sonnet` for balanced tasks, `haiku` for fast/simple tasks, or `inherit` for the parent model)

If the available context makes any of this information obvious, there's no need to ask redundantly. However, clarify any ambiguity rather than making assumptions.
If the user provided arguments, use them to inform the agent name and purpose.

### Phase 2: Design

Determine the agent's role, delegation behavior, and tooling:

**Agent vs. Skill:**
- **Agents** run as isolated subprocesses with their own context window, tool access, and model. They receive a task, work autonomously, and return a summary. Use agents for complex, multi-step work that benefits from isolation.
- **Skills** inject instructions into the current conversation's context. Use skills for workflows, conventions, or templates that should run inline.

Choose an agent when:
- The task produces verbose intermediate output that would clutter the main conversation
- The work is self-contained and can be summarized on completion
- You need to enforce specific tool restrictions or a different model
- The task benefits from a focused system prompt without main conversation noise
- You want parallel execution of independent work streams

**Nami