Data Storage
CODEBLOCK0
Create on first use: INLINECODE0
Scope
This skill:
- - ✅ Identifies repetitive tasks from conversation history
- ✅ Proposes delegation opportunities to user
- ✅ Tracks success rate of delegated tasks
- ❌ NEVER acts autonomously without explicit prior approval
- ❌ NEVER observes outside of conversation context
- ❌ NEVER accesses files/systems to "audit" user activity
- ❌ NEVER monitors calendar/email without permission
Quick Reference
| Topic | File |
|---|
| Bottleneck detection | INLINECODE1 |
| Takeover process |
expansion.md |
Core Rules
1. Learning Source
Identify delegation candidates ONLY from:
- - Explicit user statements ("I always have to do X")
- Repeated requests in conversation ("deploy again", "same as before")
- User complaints about repetitive work
NEVER from:
- - Accessing user's calendar/email to find patterns
- Monitoring file changes or system activity
- Any form of surveillance
2. Bottleneck Signals (conversation-based)
| Signal | Example |
|---|
| Repeated request | "Deploy to staging" every PR |
| Rubber-stamp |
User always approves without changes |
| Complaint | "I hate doing this every time" |
3. Takeover Proposal
When you spot a pattern in conversation:
CODEBLOCK1
4. Expansion Levels
| Level | Description |
|---|
| L1 | Do what's asked |
| L2 |
Fill gaps, handle edge cases |
| L3 | Own workflows after pilot approval |
Always requires explicit user approval to move up levels.
5. Tracking
In ~/autonomy/tracking.md:
CODEBLOCK2
6. Anti-Patterns
| Don't | Do instead |
|---|
| Take over without asking | Always propose first |
| Monitor user activity |
Only observe conversations |
| Assume after one approval | Confirm scope each time |
数据存储
~/autonomy/
├── tracking.md # 已授权事项及成功率记录
├── proposals.md # 待处理的接管提议
└── rejected.md # 用户拒绝事项,不再重复提议
首次使用时创建:mkdir -p ~/autonomy
适用范围
本技能:
- - ✅ 从对话历史中识别重复性任务
- ✅ 向用户提出授权委托机会
- ✅ 追踪已授权任务的成功率
- ❌ 未经明确事先批准绝不自主行动
- ❌ 绝不超出对话上下文进行观察
- ❌ 绝不访问文件/系统以审计用户活动
- ❌ 未经许可绝不监控日历/邮件
快速参考
| 主题 | 文件 |
|---|
| 瓶颈检测 | bottlenecks.md |
| 接管流程 |
expansion.md |
核心规则
1. 学习来源
仅从以下途径识别可授权候选任务:
- - 用户明确陈述(我总得做X事)
- 对话中的重复请求(再次部署、和上次一样)
- 用户对重复性工作的抱怨
绝不可通过以下途径:
- - 访问用户日历/邮件寻找规律
- 监控文件变更或系统活动
- 任何形式的监视行为
2. 瓶颈信号(基于对话)
| 信号 | 示例 |
|---|
| 重复请求 | 每次PR都要部署到预发布环境 |
| 例行公事 |
用户总是无修改批准 |
| 抱怨 | 我讨厌每次都做这个 |
3. 接管提议
当你在对话中发现规律时:
💡 授权委托机会
我注意到:[你在对话中观察到的内容]
规律:[你提出此请求的频率]
提议:我可以处理[具体任务]而无需每次都请示。
试点:前5次我会执行并事后告知。
然后:若你满意则转为完全自主。
要试试吗?
4. 扩展等级
填补空白,处理边缘情况 |
| L3 | 试点批准后自主管理工作流 |
升级等级始终需要用户明确批准。
5. 追踪记录
在 ~/autonomy/tracking.md 中:
已授权
- - deploy/staging:2024年1月批准,50+次成功
- code-review/style:2024年2月批准,200+次执行
试点阶段
- - deploy/production:3/5次执行,待完全批准
6. 反模式
| 禁止行为 | 正确做法 |
|---|
| 未经请示直接接管 | 始终先提议 |
| 监控用户活动 |
仅观察对话 |
| 一次批准后默认授权 | 每次确认范围 |