원클릭으로
iterative-retrieval
为 subagent、代码探索、RAG 式检索和大仓库上下文收集设计迭代检索闭环。用于初始上下文不明确、一次性读全仓会污染上下文、子代理不知道该读哪些文件、或搜索结果需要逐轮收敛时。
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
메뉴
为 subagent、代码探索、RAG 式检索和大仓库上下文收集设计迭代检索闭环。用于初始上下文不明确、一次性读全仓会污染上下文、子代理不知道该读哪些文件、或搜索结果需要逐轮收敛时。
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
| name | iterative-retrieval |
| description | 为 subagent、代码探索、RAG 式检索和大仓库上下文收集设计迭代检索闭环。用于初始上下文不明确、一次性读全仓会污染上下文、子代理不知道该读哪些文件、或搜索结果需要逐轮收敛时。 |
这个 skill 解决子代理和大仓库探索的核心问题:开始时不知道要读什么,但把所有内容都塞进上下文会让模型变差。
最多跑 3 轮,每轮都要缩小范围。
| 等级 | 判断 | 动作 |
|---|---|---|
| 高 | 直接实现目标行为、核心调用链或关键配置 | 读细节并保留路径 |
| 中 | 提供同类模式、类型、测试或接口契约 | 只抽取模式或术语 |
| 低 | 只是同名或旁路引用 | 记录后排除 |
| 无 | 与任务无关 | 加入排除条件 |
目标:[一句话说明要找什么]
范围:[允许搜索的目录/文件]
排除:[node_modules、dist、生成物、无关目录]
初始关键词:[3-8 个]
请最多执行 3 轮检索:
1. 每轮列出查询条件。
2. 对候选文件标注高/中/低/无相关性和理由。
3. 根据发现更新下一轮关键词或路径。
4. 最后只回传:关键文件、证据、缺失上下文、建议下一步。
不要回传完整 grep 输出、目录树或长日志。
## 结论
[是否已找到足够上下文]
## 关键文件
- `path/to/file`:[为什么高相关]
## 证据
- `path/to/file:line`:[可复核事实]
## 缺口
- [仍缺的文件/术语/接口]
## 下一步
- [最小可执行动作]
Stress-test a plan, design, decision, or idea through a one-question-at-a-time interview. Use when the user invokes /grill, says grill me, asks to be challenged, 拷问, 压力测试, or wants a plan stress-tested; also use when progress is blocked by an unresolved user-owned decision whose alternatives materially change the outcome after high-risk classification is ruled out. Do not use merely because work is complex, high-risk, new, or spans multiple files.
Use when the user explicitly requests Brainstorming, a design spec, Spec Gate, or the full Superpowers workflow; also use automatically for costly-to-reverse high-risk work involving architecture or service boundaries, public-contract compatibility, authentication or authorization, persistent data/schema migration, or irreversible external side effects. Do not use for ordinary behavior changes or merely because work is new, complex, or spans multiple files.
Chinese workflow for surfacing systematic fact, evidence, context, and blind-spot gaps before, during, and after agentic work. Use when the user explicitly asks for a blind-spot scan, unknown unknowns, prototypes, reference mapping, implementation notes, explainers, or quizzes; when phrases like "I'll know it when I see it" or "teach me this domain" expose unstated assumptions; or when safe progress is blocked by evidence gaps that require multiple low-cost probes. Do not use merely because work is complex, high-risk, long-running, or spans multiple files.
Use when starting or routing a non-trivial task in this workflow harness, when the user references Superpowers, or when deciding which skill, command, rule, or agent should govern the work. Establishes skill invocation discipline: check relevant process skills before acting, follow user/project instructions first, and verify before completion.
Use when you have a spec or requirements for a multi-step task, before touching code
用于功能验收、回归验证、确认修复是否生效,以及需要用真实界面、视频或日志证据判断用户流程是否正常的任务。适用于“验收一下”“测一下”“确认修复”“给我可复核证据”等请求。