最小安装顺序

第 步：Superpowers 建立完整的软件工程流程
↓ 第 步：Context7 补充当前第三方库文档
↓ 第 步：Browser 建立真实页面验收能力
↓ 第 步：GitHub Skills 接入 PR、CI 和代码审查
↓ 第 步：按项目类型补前端/产品/文档 Skill
Need not repeat entire table again due to length constraints.
---

三、按照优先级逐个介绍*

优先级 ：Superpowers——给 Agent 装上一套严格的软件工程流程 *

Please note: If you're overwhelmed by complexity or have only trivial edits。you might skip some steps of Superpowers temporarily—focus on essential ones first.
A disciplined approach reduces surprises later.
text
请使用 Superpowers 工作流完成“为使用者资料增加头像上传”功能。先通过 brainstorming 澄清需求和边界，再写实施计划；实现阶段遵循 TDD，完成后运行验证，并在宣布完成前列出测试证据和未覆盖风险。不要擅自
图片编辑、裁剪或社交分享功能。\*\*适合谁，\*\*

\

刚开始用 Agent 做真实项目的新手
\<\/li>
\

经常遇到“代码写完了但没测试”的人
\<\/li>
\<\/ul>
\*\*不适合什么？\*\*

\

修改一个错别字或替换一句话这类极小任务。<\/ul>
---

优先级 ：diagnosing-bugs——别让 Agent 靠猜修 Bug *

User Painpoint:
Agents often “patch” without evidence leading to regressions or incomplete fixes.
\
This skill forces evidence collection before action.
\
It addresses frustration when bugs appear intermittently or are hard to reproduce.
\
By keeping debugging separate from patching,you gain confidence that fixes truly solve root causes.
\
Example prompt:
text
请使用 diagnosing-bugs Skill 诊断这个问题。先复现并收集证据，不要直接修改代码。\ n 给出最小复现、候选原因、验证实验和最终根因。\ n 问题现象：点击登录后页面一直转圈 ...\`\`\`"
\ 
---
---

五、怎样查找和安装 Skill？*

If your environment provides /skill-installer。simply ask agent:
请使用 skill-installer 列出当前可安装的 curated Skills，并标注已安装与否。不必立即安装，

Select desired skills and authorize explicitly:
请安装列表中的 并告知何时可用于新任务。其实，

This approach prevents accidental installation of unknown scripts that could modify files or access sensitive APIs.
---

九、“别只装 Skill”：用 CLAUDE.md / AGENTS.md 给项目立规矩 *

The right rules file tells Agents what *always* y must follow within your repo—not just how specific tasks are performed.


Environment Type 
Recommended Filename 
Purpose 



Claude Code 
CLAUDE.md 
Project background & command restrictions 

OpenAI Codex 
AGENTS.md 
Persistent guidance at repo/dir level 




Tip: Keep both files if possible so rules apply across platforms while still allowing tool‐specific guidelines in AGENTS.md.*