Use request_user_input to resolve ambiguities. Ask up to 10 targeted questions:
Scope boundaries (in/out of scope)
Technology/architectural constraints
Priorities (critical vs nice-to-have)
Edge case handling
Success criteria
使用request_user_input工具解决歧义。最多提出10个针对性问题：
范围边界（包含/排除范围）
技术/架构约束
优先级（关键需求 vs 锦上添花的需求）
边缘情况处理
成功标准
Phase 2: Create Plan
阶段2：制定计划
Structure
结构
Overview: Brief summary and approach
Sprints: Logical phases that build on each other
Tasks: Specific, actionable items within sprints
概述：简要总结与实施方法
迭代（Sprints）：相互依赖的逻辑阶段
任务：迭代内的具体可执行项
Sprint Requirements
迭代要求
Each sprint must:
Result in demoable, runnable, testable increment
Build on prior sprint work
Include demo/verification checklist
每个迭代必须：
产出可演示、可运行、可测试的增量成果
基于前一迭代的成果构建
包含演示/验证清单
Task Requirements
任务要求
Each task must be:
Atomic and committable (small, independent)
Specific with clear inputs/outputs
Independently testable
Include file paths when relevant
Include dependencies for parallel execution
Include tests or validation method
Bad: "Implement Google OAuth"
Good:
"Add Google OAuth config to env variables"
"Install passport-google-oauth20 package"
"Create OAuth callback route in src/routes/auth.ts"
"Add Google sign-in button to login UI"
每个任务必须：
原子化且可提交（小而独立）
明确输入/输出
可独立测试
相关时包含文件路径
包含并行执行的依赖关系
包含测试或验证方法
反面示例： "Implement Google OAuth"
正面示例：
"Add Google OAuth config to env variables"
"Install passport-google-oauth20 package"
"Create OAuth callback route in src/routes/auth.ts"
"Add Google sign-in button to login UI"
Phase 3: Save
阶段3：保存
Save the file
Generate filename from request:
Extract key words
Convert to kebab-case
Add
-plan.md
suffix
Examples:
"fix xyz bug" →
xyz-bug-plan.md
"implement google auth" →
google-auth-plan.md
保存文件
根据需求生成文件名：
提取关键词
转换为短横线分隔格式（kebab-case）
添加
-plan.md
后缀
示例：
"fix xyz bug" →
xyz-bug-plan.md
"implement google auth" →
google-auth-plan.md
Phase 4: Gotchas
阶段4：注意事项
AFTER it is saved. Identify potential issues and edge cases in the plan. Address them proactively. Where could something go wrong? What about the plan is ambiguous?
Is there a missing step, dependency, or pitfall?
Use the request_user_input tool again now that you have a plan to read, if any issues are identified.
Provide the plan file location to a subagent for review, and ask it to provide feedback. Provide it useful context so it can make sound decisions. Explicitly tell it not to ask any questions. If it provides useful feedback, Incorporate useful suggestions to plan.