AI 协作项目的复盘与经验沉淀 Skill。重建项目时间线、追踪问题与解决方案、提取可复用经验。
Built from a real 2-week project: crawling and analyzing 14,139 AI projects. 24 problems, 15 of them almost forgotten.
When you work with AI on a project spanning multiple days, context breaks. Every new conversation starts with the AI having no memory of what happened. Even memory files store fragments — a decision here, a conclusion there — not the full journey. And humans forget too.
Code has Git, products have PRDs, projects have Jira. AI-collaborative work has nothing. This skill reconstructs the whole journey from scattered sources before it evaporates.
Four steps:
- Gather raw material — pull from project memory files, session memory JSONL, workspace files (scripts, docs, git log), and current conversation. Never rely on AI memory alone. Output: chronological event list with dates, problems, solutions, file changes.
- Assess and structure — group events into phases, build a numbered problem chain (root cause, solution, iteration count per problem), extract key decisions and lessons.
- Generate the retrospective document — fixed template (starting point, per-phase problem tracking tables, timeline, lessons) that adapts section names to the actual project.
- Convert experience — propose writing lessons back to project memory, creating new skills from generalizable lessons, or generating a pre-flight checklist if the project type recurs.
Full retrospective only when the project meets at least 2 of: spanned 3+ days or 5+ sessions; 3+ iterations on a major component; context broken across sessions; output has reusable value; user explicitly wants lessons preserved. Otherwise a 3-5 point lightweight summary.
| Tool | Location |
|---|---|
| TRAE | .trae/skills/project-retrospective/SKILL.md |
| Claude Code | ~/.claude/skills/project-retrospective/SKILL.md |
| QoderWork | ~/.qoderworkcn/skills/project-retrospective/SKILL.md |
Copy the project-retrospective/ folder (containing SKILL.md) into the skills directory of your tool, then trigger it with "帮我复盘一下这个项目" / "整理一下整个过程" / "踩了哪些坑" / "retrospective".
project-retrospective/
├── README.md
└── project-retrospective/
└── SKILL.md
MIT. Free to use, modify, and redistribute.
跟 AI 协作做超过一天的事，上下文一定会断。每次新开对话，AI 对你之前做过什么一无所知。就算有 memory 机制，存的也是碎片，某个决策、某条结论、某个文件路径，看不到整个项目怎么走到今天的。人自己也会忘。
写代码有 Git，做产品有 PRD，做项目有 Jira，跟 AI 协作的人什么都没有。这个 Skill 在经验蒸发之前，把散落在各处的素材重建成完整过程记录。
四步流程：
- 扒素材。从项目 memory 文件、session 级记忆文件、工作区文件（脚本、文档、git log）、当前对话上下文四个来源系统拉取，不靠 AI 回忆。产出按时间排序的事件清单，每条带日期、发生了什么、遇到什么问题、怎么解决的、动了哪些文件。
- 找问题链。按阶段分组，给每个问题编号，记根因、解决方案、迭代次数，提取关键决策点和教训。
- 生成文档。套用固定模板，起点、各阶段过程（含问题追踪表）、时间线、核心教训，章节名随实际项目调整。
- 经验转化。文档生成完不是结束，主动建议把新教训写回项目 memory、把通用经验做成新 Skill、项目类型会复现则生成预检清单。
只有满足以下条件中至少两个，才启动完整复盘：跨度超过三天或五个 session、核心环节改过三个版本以上、中途换任务导致上下文断裂、产出物有复用价值、用户明确想保留经验。几小时的一次性任务只给三到五条要点摘要。
| 工具 | 位置 |
|---|---|
| TRAE | .trae/skills/project-retrospective/SKILL.md |
| Claude Code | ~/.claude/skills/project-retrospective/SKILL.md |
| QoderWork | ~/.qoderworkcn/skills/project-retrospective/SKILL.md |
把含 SKILL.md 的 project-retrospective/ 文件夹拷进对应工具的 skills 目录，然后对 AI 说"帮我复盘一下这个项目"、"整理一下整个过程"、"我们之前踩了哪些坑"，或者直接说"retrospective"。
project-retrospective/
├── README.md
└── project-retrospective/
└── SKILL.md
MIT，随便用，随便改，随便传。