Let the Agent learn from mistakes and continuously improve
Quick start
用户: 记录一下刚才的问题
用户: 更新口误识别的规则
用户: 这个教训要记下来
Update location
| Content type | Target file | Example |
|---|---|---|
| User profile | CLAUDE.md |
Preferences, habits |
| Methodology + feedback | */tips/*.md |
Rules, lessons |
Process
用户触发("刚才失败了"、"记录一下")
↓
【自动】回溯上下文,找出问题点
↓
【自动】读目标文件全文,理解现有结构
↓
【自动】整合到正文相应位置(不是只往末尾加!)
↓
【自动】反馈记录只记事件,不重复规则
↓
汇报更新结果
Key: Don't ask "what question", analyze directly from the context!
Update principles
❌ Wrong approach: append to the end
## 反馈记录
### 2026-01-14
- 教训:审查稿末尾必须生成删除任务清单
- 教训:用户确认时要分别确认口误和静音
Only adding to the feedback record = rules scattered at the end, and the same mistake will happen again next time
✅ Correct approach: integrate into the main body
- Read the full text, understand the chapter structure
- Find the corresponding location, integrate the rules into it
- The feedback record only logs events:
- 审查稿标记了静音,但剪辑时漏删
## 四、审查稿格式
(新增删除任务清单模板)
## 五、确认与执行流程 ← 缺这个章节就新增
(新增分别确认口误和静音的流程)
## 反馈记录
### 2026-01-14
- 审查稿标记了静音,但剪辑时漏删(只删了口误)
Trigger conditions
- User corrects an AI mistake
- The user says "remember this", "pay attention from now on"
- Discover a new general pattern
Counterexample
2026-01-13
❌ 错误:
用户: 刚才失败了,更新到skills
AI: 请告诉我你发现了什么问题? ← 不该问!
✅ 正确:
AI: [自动回溯上下文,找到失败点]
AI: [执行更新]
2026-01-14
❌ 错误:
AI: 已更新,在反馈记录新增3条教训 ← 只加末尾!
✅ 正确:
AI: [读全文,理解结构]
AI: [整合到正文相应位置]
AI: [反馈记录只记事件]
AI: 已更新:新增第五章"确认与执行流程",更新第四章模板
Principle: Rules should be integrated into the main body; the feedback record is only an event log