Create Agent Skill

Generate a new custom agent .md file in ~/.claude/agents/ that matches the user's existing agent conventions.

Workflow

1. Gather Requirements

Ask the user for the following (skip any they already provided):

• name: kebab-case agent name (e.g., qa-tester, tech-writer)
• role: What this agent does (1-2 sentences)
• model: haiku (lightweight tasks), sonnet (development/analysis), or opus (complex reasoning/strategy)

2. Analyze Existing Agents

Read 2-3 existing agents from ~/.claude/agents/ to match the current conventions:

• Frontmatter format (name, description, model, memory, tools)
• Section structure (role intro → responsibilities → expertise → workflow → collaboration → communication)
• Tone and detail level

3. Generate Agent File

Create ~/.claude/agents/{name}.md with this structure:

---
name: {name}
description: "{1-2 sentence description with 3-5 usage examples}"
model: {model}
memory: user
---

{Role introduction paragraph - 1-2 sentences, senior professional persona}

## Core Responsibilities

{4-6 numbered responsibilities}

## Technical Expertise (or domain-specific equivalent)

{Relevant skills, tools, frameworks organized by category}

## Workflow

{Step-by-step working process}

## Collaboration

- Work with relevant existing agents (ceo, cso, planner, frontend-dev, backend-dev, mobile-dev, ai-engineer, data-engineer, devops, researcher, reviewer, writer, doc-translator, git-master)
- Submit work to **reviewer** for quality gate (if applicable)
- Follow **planner**'s task assignments (if applicable)

## Communication

- Respond in user's language
- Use `uv run python` for Python execution

**Update your agent memory** as you discover {domain-specific learnings}.

4. Configuration Decisions

• tools: Only restrict if the agent should NOT have full tool access (e.g., read-only agents). Default: omit (grants all tools).
• model: Match complexity to cost:
  • haiku: Translation, git ops, simple formatting (Haiku 4.5)
  • sonnet: Code writing, analysis, debugging (Sonnet 4.6)
  • opus: Strategy, complex reasoning, architecture decisions (Opus 4.6)

5. Verify

• Confirm the file was created at the correct path
• Show the user the complete agent file for review

Rules

• Match the style and depth of existing agents — don't be more verbose or more terse
• Description examples in frontmatter should use the arrow format: "Do X" → Launch {name}
• Always include memory: user in frontmatter
• Korean persona descriptions are fine if user speaks Korean, but agent names must be English kebab-case
• Don't add unnecessary sections — keep it focused on the agent's actual domain