agent-builder

This skill helps create specialized Claude Code agents adapted to a specific project context.

When to use

• Starting a project with specific expertise needs
• Need a "creative director" or domain expert
• Existing agents (infra-expert, creative-director) don't cover the domain
• Want to capitalize expertise for a recurring project

Process

1. Understand the need

Ask these questions:

Domain

• What is the area of expertise? (frontend, backend, UX, creative, business...)
• Which specific frameworks/tools?

Stance

• Review only (read-only) or can modify code?
• Style: directive, socratic, provocative?

Focus

• What are the 3-5 critical points to always check?
• What are common anti-patterns to detect?

Project context

• One-time or recurring project?
• Solo or team collaboration?

2. Choose the template

Two templates available:

• expert-template.md — For technical experts (review, audit, advice)
• creative-template.md — For creative roles (brainstorm, direction, ideation)

3. Generate the agent

Use the appropriate template and customize:

• Name and description (for frontmatter)
• Allowed tools
• Domain-specific checklist
• Adapted response format

4. File placement

• User-level (~/.claude/agents/): if reusable across projects
• Project-level (.claude/agents/): if specific to this project

Available templates

See templates/ folder:

• expert-template.md: structure for technical expertise agents
• creative-template.md: structure for creative/strategic agents

Example agents to create

Frontend Expert

• Domain: React, TypeScript, CSS
• Focus: accessibility, performance, reusable components
• Tools: Read, Grep, Glob, Bash(npm:, eslint:)

Creative Director

• Domain: product ideation, UX strategy
• Stance: provocative, challenges assumptions
• Tools: Read (read-only, no code)

Domain Expert (e.g., Finance)

• Domain: specific business rules
• Focus: compliance, business edge cases
• Tools: Read, Grep

Best practices

1. Keep checklists short — 10-15 points max, otherwise dilution
2. Be specific — "check React hooks" > "check code"
3. Include anti-patterns — what NOT to do
4. Define output format — structure the agent's responses
5. Version with project — if project-level, commit the agent