home / skills / ceeon / videocut-skills / 自更新

自更新 skill

safe

/自更新

This skill records user feedback, updates rules and methodologies, and keeps documentation current across CLAUDE.md and tips directories.

npx playbooks add skill ceeon/videocut-skills --skill 自更新

Review the files below or copy the command above to add this skill to your agents.

Files (2)

SKILL.md

2.7 KB

---
name: videocut:自更新
description: 自更新 skills。记录用户反馈，更新方法论和规则。触发词：更新规则、记录反馈、改进skill
---

<!--
input: 用户反馈、错误纠正
output: 更新后的文档（CLAUDE.md 或 tips/*.md）
pos: 元 skill，让 Agent 从错误中学习

架构守护者：一旦我被修改，请同步更新：
1. ../README.md 的 Skill 清单
2. /CLAUDE.md 路由表
-->

# 自更新

> 让 Agent 从错误中学习，持续改进

## 快速使用

```
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
```

## 更新位置

| 内容类型 | 目标文件 | 示例 |
|---------|---------|------|
| 用户画像 | `CLAUDE.md` | 偏好、习惯 |
| 方法论 + 反馈 | `*/tips/*.md` | 规则、教训 |

## 流程

```
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
```

**关键**：不要问"什么问题"，直接从上下文分析！

## 更新原则

### ❌ 错误做法：往末尾加

```markdown
## 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
```

只加到反馈记录 = 规则散落在末尾，下次还会犯错

### ✅ 正确做法：整合到正文

1. **读全文**，理解章节结构
2. **找到相应位置**，把规则整合进去
3. **反馈记录只记事件**：`- 审查稿标记了静音，但剪辑时漏删`

```markdown
## 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）
```

## 触发条件

- 用户纠正 AI 错误
- 用户说"记住这个"、"以后注意"
- 发现新的通用规律

## 反例

### 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
```

**原则**：规则要整合到正文，反馈记录只是事件日志

Overview

This skill enables an agent to self-update by recording user feedback and integrating lessons into its documentation. It captures error events, derives corrective rules, and merges them into the right sections of existing guidance files so the agent learns from mistakes. It reports what changed and where.

How this skill works

When triggered by user feedback or detected failures, the agent automatically backtracks recent context to locate the failure point. It reads the target documentation (e.g., CLAUDE.md and tips/*.md), determines the correct insertion points, and integrates new rules or templates into the proper sections rather than appending notes to the end. Finally, it logs a concise event record and summarizes the applied updates.

When to use it

User corrects an AI mistake or says 'remember this' or 'watch out next time'.
A recurring error pattern is observed across edits or outputs.
You want rules and templates added to existing methodology sections.
To convert incident details into actionable process changes.
When maintaining living documentation and workflow guidance.

Best practices

Automatically analyze recent context; do not ask the user to restate the problem.
Always read the full target document to understand structure before editing.
Integrate new rules into the relevant section; avoid piling updates at the end.
Log only event details in feedback records; keep rules inside the methodology.
Report concrete changes: which section was added or updated and why.

Example use cases

User reports missed mute segments; agent inserts a 'confirm mute' checklist into the execution flow and logs the incident.
User flags repeated transcription mistakes; agent updates 'speech-to-text' rules in tips and records the failure event.
Agent detects a process gap; it creates a new chapter for 'confirmation and execution' and notes the triggering event.
User asks the agent to 'record what went wrong'; agent backtracks, extracts the issue, integrates a fix, and summarizes updates.

FAQ

What triggers a self-update?

User corrections, explicit prompts like 'remember this', or detected recurring errors will trigger automatic analysis and updates.

Will the agent ask the user to describe the problem?

No. The agent should automatically backtrack the context and identify the failure point without asking the user to repeat details.