菜单
Use before building any non-trivial feature — when a prompt would otherwise make the agent guess at data models, API shape, auth, state, or module boundaries that haven't been decided yet. Use when you catch yourself about to "just start coding" something with real structure.
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
Use when setting up a project's structure, or when the agent needs the stack, folder layout, data flow, schemas, and the hard boundary rules it must never cross. Use when code keeps crossing layers (UI logic in API routes, DB calls in components).
Use when planning what to build and in what order — splitting a project into sequenced, scoped feature phases so the agent never has to decide "what's next" on its own. Use when work feels like random feature-hopping.
Use when establishing or enforcing how code is written in a project — naming, conventions, file structure, error handling — so the agent stays consistent across sessions instead of drifting into its own style by week two.
Use when a project depends on third-party libraries and the agent needs project-specific usage rules — and must check for an MCP server or current docs — before coding against a library it only half-remembers.
Use when tracking build progress across sessions — a living checklist updated as features complete, so a new session knows exactly where work left off without re-explaining.
Use when starting a new project, or when the agent lacks a clear picture of what's being built, who it's for, and what is explicitly out of scope. Use when a build keeps producing things nobody asked for.
| name | architect |
| description | Use before building any non-trivial feature — when a prompt would otherwise make the agent guess at data models, API shape, auth, state, or module boundaries that haven't been decided yet. Use when you catch yourself about to "just start coding" something with real structure. |
The most expensive bug is a structural decision made by guessing mid-build. architect forces the open decisions into the light before any code, so the agent builds against a plan instead of inventing one halfway through.
Run it in plan mode (no edits) before any complex feature.
Skip for: pure styling, copy changes, a one-file utility — anything with no real design decision.
A plan the human approves before implementation begins. This plan IS the contract the later review step checks against.
