你已安装Superpowers，准备开始实际使用。这篇文章通过一个完整的Todo应用案例，展示Superpowers的每个阶段。
我们不是看概念，而是看真实的对话和行动。
场景设定：我们要构建什么
目标：一个简单的Todo应用。
功能：
- 添加任务
- 删除任务
- 标记任务完成
技术栈：Node.js + Express + 内存存储（简单起见，不用数据库）
阶段一：Brainstorming（需求探索）
我启动Claude Code，输入：
我想创建一个Todo应用
AI没有直接开始写代码。它说：
I notice this is about creating a feature. I need to use the brainstorming skill first.
Let me understand your project context. I'll check files and recent commits.
AI读取了当前目录（空的），然后开始问问题：
Before we start designing, I have some questions:
1. What is this project for?
- Personal use?
- Part of a larger application?
- A demo/learning project?
2. Who will use this Todo app?
- Just you?
- A team?
- End users of a product?
3. What are the success criteria?
- Basic CRUD functionality?
- Specific performance requirements?
- Integration with other systems?
这是brainstorming的核心：AI不猜测，而是问。
我回答：
1. A demo project to learn Superpowers workflow
2. Just me for testing
3. Basic CRUD: add, delete, mark complete. No extra features.
AI继续问：
Good. A few more questions:
4. What's the API style?
- REST API?
- CLI interface?
- Web UI?
5. How should tasks be stored?
- In-memory (simple, but lost on restart)?
- File-based (persists)?
- Database?
6. Any specific format for tasks?
- Just text?
- Priority levels?
- Due dates?
我回答：
4. REST API (I want to test with HTTP requests)
5. In-memory is fine for this demo
6. Just text. No priority, no dates. Keep it simple.
AI形成了一个初步设计：
Based on our conversation, I'll present the design in sections.
## Architecture
- Node.js + Express
- REST API endpoints:
- POST /todos - create task
- GET /todos - list all tasks
- DELETE /todos/:id - delete task
- PUT /todos/:id/complete - mark complete
- In-memory storage: array of task objects
- Task object: { id, text, completed }
Does this architecture look right so far?
我说：
Yes, continue.
AI继续：
## Components
- server.js: Express app and route definitions
- todoStore.js: In-memory storage and CRUD operations
- routes/todos.js: Route handlers (optional, can be in server.js for simplicity)
## Testing
- Jest for testing
- Test coverage for all CRUD operations
- Integration tests for API endpoints
Does this look right?
我说：
Yes, but keep routes in server.js for simplicity.
AI调整并继续：
Adjusted. One file: server.js contains all.
## Error Handling
- 400 for invalid input (empty text)
- 404 for non-existent task ID
- 500 for unexpected errors
Does this look right?
我说：
Yes.
AI完成设计：
Design complete. I'll write the spec document.
Writing to docs/superpowers/specs/2026-04-28-todo-app-design.md...
AI写了一个spec文件，然后进行自检：
Spec written. Let me review it for placeholders, contradictions, and ambiguity.
Self-review passed. No placeholders, internally consistent, scope focused, no ambiguity.
Please review the spec file and let me know if you want to make changes before we proceed.
我打开spec文件，确认内容正确，回复：
Looks good. Proceed.
阶段二：Writing Plans（实现计划）
AI说：
I'll use the writing-plans skill to create an implementation plan.
Breaking the spec into bite-sized tasks (each 2-5 minutes)...
AI生成了计划：
# Implementation Plan
## Task 1: Setup Project
- Create package.json with Express and Jest
- Create server.js skeleton
- Verify: npm install succeeds, npm test runs (no tests yet)
## Task 2: Implement TodoStore
- Write test for addTodo
- Watch test fail
- Implement addTodo function
- Watch test pass
- Write test for deleteTodo
- Watch test fail
- Implement deleteTodo
- Watch test pass
- Write test for getTodos
- Watch test fail
- Implement getTodos
- Watch test pass
- Write test for markComplete
- Watch test fail
- Implement markComplete
- Watch test pass
- Verify: all tests pass
## Task 3: Implement API Endpoints
- Write test for POST /todos
- Watch test fail
- Implement POST /todos endpoint
- Watch test pass
- Write test for GET /todos
- Watch test fail
- Implement GET /todos endpoint
- Watch test pass
- Write test for DELETE /todos/:id
- Watch test fail
- Implement DELETE /todos/:id endpoint
- Watch test pass
- Write test for PUT /todos/:id/complete
- Watch test fail
- Implement PUT /todos/:id/complete endpoint
- Watch test pass
- Verify: all API tests pass
## Task 4: Error Handling
- Write test for POST with empty text (expect 400)
- Watch test fail
- Add validation for empty text
- Watch test pass
- Write test for DELETE non-existent ID (expect 404)
- Watch test fail
- Add 404 handling
- Watch test pass
- Verify: all error tests pass
## Task 5: Final Verification
- Run all tests
- Verify: 100% coverage
- Manual test with curl
- Verify: all endpoints work
AI问：
Plan created. Does this look good?
我说：
Yes, proceed.
阶段三：Worktree（隔离工作空间）
AI说：
I'll use the using-git-worktrees skill to create an isolated workspace.
Creating worktree on a new branch...
AI执行：
git worktree add .claude/worktrees/todo-app -b todo-app
然后：
Worktree created at .claude/worktrees/todo-app on branch todo-app.
Setting up project in worktree...
Verifying clean test baseline (no tests yet, baseline clean)...
You can observe my work in the worktree directory, or I can report here.
我选择观察，但主要让AI自主工作。
阶段四：Subagent-driven Development（执行）
AI说：
I'll use the subagent-driven-development skill to execute the plan.
Each task will be handled by an independent subagent with review loops.
任务1：Setup Project
AI dispatch子代理：
Dispatching subagent for Task 1: Setup Project...
Subagent report:
- Created package.json with Express and Jest
- npm install succeeded
- Created empty server.js
- npm test runs (no tests, baseline clean)
Self-review: Completed all items in task spec.
审查者（另一个子代理）检查：
Spec compliance review:
- package.json created ✓
- Express and Jest dependencies ✓
- server.js skeleton ✓
- npm install succeeds ✓
- npm test runs ✓
No issues. Proceed to code quality review.
Code quality review:
- package.json format correct ✓
- server.js minimal skeleton ✓
No issues. Task 1 complete.
任务2：Implement TodoStore
这个任务展示了TDD流程：
Dispatching subagent for Task 2: Implement TodoStore...
Subagent report (detailed):
RED phase - addTodo:
- Wrote test: test('addTodo creates task with id')
- Ran test: FAIL - addTodo not defined
- Failure message correct: "ReferenceError: addTodo is not defined"
- Fails because feature missing ✓
GREEN phase - addTodo:
- Implemented minimal addTodo function in todoStore.js
- Ran test: PASS
REFACTOR: Not needed, minimal implementation
RED phase - deleteTodo:
- Wrote test: test('deleteTodo removes task')
- Ran test: FAIL - deleteTodo not defined
- Failure message correct ✓
GREEN phase - deleteTodo:
- Implemented deleteTodo
- Ran test: PASS
... (similar for getTodos and markComplete)
Self-review: All tests pass. Coverage: 100% for todoStore.js.
审查者检查：
Spec compliance review:
- Tests written before code ✓
- Tests watched failing ✓
- Minimal implementations ✓
- All tests pass ✓
Checking TDD compliance for each function:
- addTodo: test first, fail verified, code minimal ✓
- deleteTodo: test first, fail verified, code minimal ✓
- getTodos: test first, fail verified, code minimal ✓
- markComplete: test first, fail verified, code minimal ✓
No issues.
Code quality review:
- Code clear and minimal ✓
- No extra features added ✓
- Tests cover edge cases ✓
No issues. Task 2 complete.
我观察到：AI严格执行TDD。每个函数都先写测试，运行看失败，再写代码。
任务3和4：API和错误处理
类似流程，每个endpoint都经过TDD循环。
审查者检查spec compliance：确认实现了所有endpoint，确认错误处理正确。
任务5：Final Verification
Dispatching subagent for Task 5: Final Verification...
Subagent report:
- All tests pass (15 tests)
- Coverage: 100%
- Manual test with curl:
- POST /todos: created task with id 1 ✓
- GET /todos: returned [ { id: 1, text: "test", completed: false } ] ✓
- PUT /todos/1/complete: returned updated task ✓
- DELETE /todos/1: returned 204 ✓
- GET /todos: returned [] ✓
All verification passed.
阶段五：Finishing（完成）
AI说：
All tasks complete. Using finishing-a-development-branch skill.
Verifying tests: all pass.
Verifying git status: all changes committed.
Options:
1. Merge to main
2. Create PR
3. Keep branch for later
4. Discard work
What would you like to do?
我选择：
Create PR.
AI执行：
Pushing branch todo-app to origin...
Creating PR...
PR URL: https://github.com/your-repo/pull/1
Cleaning up worktree...
AI删除了worktree，保持工作目录干净。
我观察到的AI行为
整个过程大约50分钟。我的参与主要在：
- Brainstorming阶段：回答问题，批准spec
- Plan阶段：批准计划
- PR阶段：选择创建PR
其余时间AI自主工作。我看到：
- AI严格执行TDD：没有一次跳过测试
- AI自我检查：每个子代理报告前都自检
- 审查循环：审查者独立验证，不信任实现者报告
- 没有添加额外功能：只做了spec里定义的
时间对比
传统方式（没有Superpowers）：
- 直接写代码：10分钟
- 手动测试：5分钟
- 发现bug修复：5分钟
- 添加测试（后补）：10分钟
- 总计：30分钟，但不确定覆盖
Superpowers方式：
- Brainstorming：10分钟
- Plan：5分钟
- Execution：35分钟
- 总计：50分钟，但确定覆盖
多花了20分钟，但结果是：
- 有spec文档，可以回顾
- 有完整测试，覆盖率100%
- 没有猜测，知道每个函数为什么存在
- 可以安全重构
总结
Superpowers的完整流程：
- Brainstorming：问答探索需求，形成spec，你批准
- Writing Plans：分解为任务，你批准
- Worktree：隔离环境，干净基线
- Subagent-driven Development：AI自主执行，TDD循环，审查验证
- Finishing：验证测试，选择merge/PR/discard
你的角色：参与关键决策点（设计批准、计划批准、PR选择），观察AI自主工作。
AI的角色：执行方法论，严格执行TDD，自我检查，独立审查。
下一步
下一篇《TDD实战：AI助手的测试驱动开发强制执行》将深入Superpowers的TDD技能：
- RED-GREEN-REFACTOR流程详解
- 为什么"看测试失败"是必要的
- AI如何抵抗"跳过测试"的压力
