# Agent Team — Reference Templates ## Overview These are **reference templates** that the `/init-team` skill adapts for each project. Each file defines an agent archetype with personality, role boundaries, and tool access. When `/init-team` runs, it reads these templates and rewrites them with project-specific context — technology stack, domain vocabulary, file paths, and team structure. Do not use these templates directly. Run `/init-team` to generate a project-tailored agent team. ## Tier System ### Development Tier (Precise) Agents that write, test, and maintain code. Methodical, exact, structured. | Agent | Style | Model | Focus | |-------|-------|-------|-------| | `architect` | Precise + Holistic | opus | System design, feasibility, tech decisions | | `developer-backend` | Precise + Organized | sonnet | Server-side code, APIs, data models | | `developer-frontend` | Precise + Creative | sonnet | UI implementation, interaction, accessibility | | `qa-engineer` | Precise + Organized | sonnet | Testing, verification, edge cases | ### Creative Tier (Expressive) Agents that shape what the project looks, sounds, and reads like. | Agent | Style | Model | Focus | |-------|-------|-------|-------| | `visual-designer` | Creative + Empathic | sonnet | Art direction, UI patterns, visual coherence | | `audio-designer` | Creative + Empathic | sonnet | Sound design, audio architecture | | `content-author` | Creative + Precise | sonnet | Copy, in-app text, voice consistency | ### Operations Tier (Organized) Agents that keep the project moving and documented. | Agent | Style | Model | Focus | |-------|-------|-------|-------| | `project-manager` | Organized + Holistic | sonnet | Tickets, sprints, coordination | | `librarian` | Organized + Precise | sonnet | Documentation, decisions, search | ### Design Tier (Holistic) Agents that shape how systems fit together and whether they serve the user. | Agent | Style | Model | Focus | |-------|-------|-------|-------| | `designer` | Holistic + Creative | sonnet | Systems design, UX, feature evaluation | | `consultant` | Holistic + Precise | sonnet | Devil's advocate, stress-testing, benchmarks | ### Stakeholder Panel (Empathic, Read-Only) Simulated user perspectives. These agents cannot write code — they observe, react, and advocate. | Agent | Style | Model | Focus | |-------|-------|-------|-------| | `power-user` | Empathic | sonnet | Depth, configurability, expert features | | `casual-user` | Empathic | sonnet | Simplicity, discoverability, onboarding | | `product-advocate` | Empathic | sonnet | ROI, prioritization, value delivery | | `marketer` | Empathic | sonnet | Positioning, messaging, demo moments | ## Usage Modes ### Single session (brainstorming) Spawn one or two agents into a conversation for focused feedback. Good for design reviews, architecture discussions, or content critique. ### Subagent delegation (focused tasks) Launch an agent via the Task tool for a specific, scoped piece of work — write a test suite, review a PR, document a decision. The agent works autonomously and returns results. ### Agent teams (parallel work) Create a team with TeamCreate, assign tasks, and let multiple agents work in parallel. Best for sprints, large refactors, or content production runs. ## How to Add a New Agent 1. Create a `.md` file in `.claude/agents/` with YAML frontmatter 2. Define personality, role boundaries, and context instructions 3. Create a briefing at `docs/briefings/{name}.md` 4. Add the agent to `TEAM.md` 5. Run `/init-team` to regenerate project-specific versions ## Model Recommendations - **opus**: Reserve for the architect. Complex multi-system reasoning, technology evaluation, architecture decisions that affect the entire project. - **sonnet**: Default for all other agents. Strong enough for implementation, creative work, and coordination. - **haiku**: Quick lookups, simple formatting tasks, status checks. Use when speed matters more than depth. ## Tool Access Patterns - **Full access** (Read, Glob, Grep, Edit, Write, Bash): Development and Operations agents - **Write access** (Read, Glob, Grep, Edit, Write): Creative agents that produce content but don't run commands - **Read-only** (Read, Glob, Grep): Stakeholder panel — they observe and comment, never modify - **Extended** (+ WebSearch, WebFetch): Architect and Consultant for research tasks