菜单
Scaffold a brand-new team under `teams/<slug>/` for the claude-agent-teams CLI. Gathers the goal from the user, sketches the component list, then delegates to the `team-author` subagent to generate files and verify install round-trip. Use when the user says "create a team for X" or "add a new team".
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Architect produces a full web-app system design from the approved product vision and prototype (the **Design** phase of the spec-driven loop) — extracted architecture drivers, deep exploration of each load-bearing (one-way-door) decision weighed against those drivers, ADRs (architecture style, framework, DB, auth, rendering strategy, multi-tenancy), C4 context/container diagrams, data model, API contracts, component breakdown, observability + security plan, technical verification criteria that bind back to the vision's product criteria, and a phased implementation plan. Use after product vision and prototype are approved.
Architect decomposes the approved web-app system design into milestones and INVEST-sized tasks, splitting each feature slice into a backend task (API + business logic + data, test-driven) and a dependent frontend task (the UI that consumes it), with acceptance criteria, dependencies, and a parallelization plan (the **Tasks** phase of the spec-driven loop). Each task declares its discipline and which system-design verification criteria it advances, so the spec→implementation lineage is explicit. Starts with a walking skeleton (real, deployable, end-to-end). Use after system design is approved.
CEO revisits the product vision and knowledge base — re-reads everything, checks what's changed, talks to the client, and updates the strategic documents. Use when priorities shift, scope changes, or the CEO needs a full refresh.
DBA designs the database schema for a web application from the system design — tables, indexes, constraints, RLS, zero-downtime migration plan. Defaults to Postgres + Redis. Produces .claude/database-schema.md. Use after system design is approved.
Designer extracts a web design specification from the approved HTML prototype — design tokens (CSS custom properties for color/type/spacing/radius/shadow), component inventory, screen map with visual acceptance criteria per screen, responsive behavior, and interaction states. This becomes the single source of truth for UI implementation. Use after the prototype is approved.
Project kickoff — CEO has a natural conversation with the client, plays devil's advocate, crystallizes the product vision (the **Specify** phase of the spec-driven loop) with explicit verification criteria, then iterates with the designer on visuals until approved. No implementation details — pure product thinking. Use at the very start of a new project.
| name | team-create |
| description | Scaffold a brand-new team under `teams/<slug>/` for the claude-agent-teams CLI. Gathers the goal from the user, sketches the component list, then delegates to the `team-author` subagent to generate files and verify install round-trip. Use when the user says "create a team for X" or "add a new team". |
| user-invocable | true |
| allowed-tools | Read, Write, Edit, Glob, Grep, Bash, Agent |
| argument-hint | [<slug>] [<one-line goal>] |
Create a new team in teams/<slug>/ that the claude-agent-teams CLI can install cleanly.
Parse $ARGUMENTS:
If either is missing, ask the user — one sharp question, not a survey:
"What should the team be called (slug), and what's the one-line goal?"
If the slug already exists at teams/<slug>/, stop and tell the user. Suggest using /team-edit instead, or pick a different slug.
Before writing anything, think through what this team actually needs. Propose to the user a one-bullet-per-file list:
teams/<slug>/
- team.json
- agents/<role>.md (only if the team needs delegated roles)
- skills/<slug>-<verb>/SKILL.md (only if a slash-command entry point makes sense)
- hooks/hooks.json + scripts/... (only if invariants must be enforced)
- .mcp.json (only if an MCP server is genuinely needed)
Keep it minimal. A team with one purposeful agent is fine. Get a thumbs-up from the user before generating.
team-authorSend the team-author subagent with:
CLAUDE.md and the bundled teams as the style reference.Let the subagent produce the files, chmod +x any scripts, and run the install round-trip test in a temp dir.
Summarize for the user in a few bullets:
/team-review it before committing