Yueyanc

Self-Improving Proactive Agent

AI 智能体v1.0.0
作者:Yueyanc更新于 2026-05-11self-improving-proactive-agent@clawhub_yueyanc/self-improving-proactive-agent
任务自动化工作流编排上下文管理#Self Improving#Self Improvement#Learn

据 AlphaMoat 统计,截至 2026-10,Self-Improving Proactive Agent 累计下载 34,965 次、获 14 个星标,按累计下载量在本次参与排名的 10,914 个 AI 技能中排名第 766 位,在「AI 智能体」分类(1,632 个)中排名第 127 位。

列表默认按累计下载量排序,也可按星标排序;详情页排名按累计下载量计算。

累计下载
3.5万
星标
14
全站下载量排名
#766
全站前 8% · 「AI 智能体」前 8%
平台累计下载星标评论
ClawHub ↗1.8万11-
SkillHub ↗3.5万14—

各平台的计数口径不同,统计时间不一致。

这个技能是做什么的

技能分两层:一层从用户纠正、复盘与反复成功中学习并维护长期状态,另一层维持推进节奏、快速恢复上下文并主动给出下一步。它在范围里写明从不在记忆文件中存放凭据或密钥,也不会在未获批准时发消息、花钱或删除数据。 建议只是提案,不保证必然改善;请检查持久状态,且不得保存凭证或密钥。

资料发布于 2026-10-29

可以用来做什么

  • 为多步任务加上纠错学习与上下文恢复机制
  • 整理一套主动推进的边界规则,避免越权动作

什么时候用

  • 用户纠正过行为或表达过长期偏好时
  • 任务容易中途跑偏,需要恢复上下文并继续推进时

示例提示词

以下是根据技能文档生成的示例,不代表实际运行结果。

按这套规则设计记忆的晋级与衰减策略,并说明哪些动作必须先征求同意
帮我把「纠错后要记住什么」写成一份可执行的清单

月度数据趋势

累计下载3.5万
8,1432.3万3.7万2026-03 · 累计下载 1.0万2026-05 · 累计下载 1.5万2026-06 · 累计下载 1.5万2026-07 · 累计下载 1.6万2026-08 · 累计下载 1.8万2026-09 · 累计下载 2.0万2026-10 · 累计下载 3.5万26-0326-0926-10
星标14
611152026-03 · 星标 72026-05 · 星标 142026-06 · 星标 142026-07 · 星标 142026-08 · 星标 142026-09 · 星标 142026-10 · 星标 1426-0326-0926-10

2026-09 起统计方法调整:同名技能按作者分别统计,并覆盖全部技能。图中虚线为调整前的月份,与之后的数据不直接比较,也不计算跨越调整的增长。 该技能有同名的其他版本,调整前的数据归属当前下载量最大的这一个。

月份累计下载环比星标
2026-103.5万—14
2026-092.0万—14
2026-081.8万+7.6%14
2026-071.6万+6.3%14
2026-061.5万+4.6%14
2026-051.5万—14
2026-031.0万—7

SKILL.md 原文

以下是作者发布的原始文件,按纯文本显示,本站未做修改或审核;其中的链接不可点击,命令请核实后再运行。抓取于 2026-10-23。

6.6 KB · ClawHub v1.0.0
---
name: Self-Improving Proactive Agent
slug: self-improving-proactive-agent
version: 1.0.0
homepage: https://github.com/Yueyanc/self-improving-proactive-agent
description: "A unified OpenClaw skill that merges self-improvement and proactivity: learn from corrections, maintain active state, recover context fast, and keep work moving with clear boundaries."
changelog: "Initial release. Combines the strongest patterns from self-improving and proactivity into one canonical skill package."
metadata: {"clawdbot":{"emoji":"🧠","requires":{"bins":[]},"os":["linux","darwin","win32"],"configPaths":["~/self-improving/","~/proactivity/"],"configPaths.optional":["./AGENTS.md","./SOUL.md","./HEARTBEAT.md","./TOOLS.md"]}}
---
展开全部(共 214 行)
# Self-Improving Proactive Agent

One skill, two layers:

- **Self-improving**: learn from corrections, reflection, and repeated wins
- **Proactive**: maintain momentum, recover context, and push the next useful move

Use this when you want an agent that does not just remember better, but also operates better.

## When to Use

Use this skill when:
- the user corrects you or states durable preferences
- the task is multi-step or likely to drift
- context recovery matters
- follow-through and heartbeat behavior should improve over time
- the user wants a single unified behavior model instead of separate overlapping skills

## Unified Architecture

```text
~/self-improving/
├── memory.md               # HOT: confirmed durable rules and preferences
├── corrections.md          # recent corrections and reusable lessons
├── index.md                # storage map / topic index
├── heartbeat-state.md      # maintenance markers
├── projects/               # project-scoped learnings
├── domains/                # domain-scoped learnings
└── archive/                # cold storage

~/proactivity/
├── memory.md               # stable activation and boundary rules
├── session-state.md        # current objective, decision, blocker, next move
├── heartbeat.md            # lightweight recurring follow-through
├── patterns.md             # reusable proactive wins
├── log.md                  # recent proactive actions
└── memory/
    └── working-buffer.md   # volatile breadcrumbs for long / fragile tasks
```

## Core Principles

### 1. Learn from explicit evidence
Learn from:
- direct user corrections
- explicit preferences
- repeated successful workflows
- self-reflection after meaningful work

Do not learn from:
- silence
- vibes alone
- one-off context instructions
- unverified assumptions

### 2. Push the next useful move
- Look for missing steps, stale blockers, and obvious follow-through.
- Prefer drafts, checks, patches, and prepared options.
- Stay quiet when the value is weak.

### 3. Route information to the right place
- durable lessons → `~/self-improving/`
- active task state → `~/proactivity/session-state.md`
- volatile breadcrumbs → `~/proactivity/memory/working-buffer.md`

### 4. Recover before asking
Before asking the user to restate work:
1. read HOT self-improving memory
2. read proactive stable memory
3. read session state
4. read working buffer when needed
5. ask only for the missing delta

### 5. Verify implementation, not intent
If you changed how something works:
- change the real mechanism, not just wording
- test the outcome from the user perspective
- only then report success

### 6. Stay proactive inside hard boundaries
Always ask first for:
- messages or contact
- spending money
- deleting data
- public actions
- commitments or scheduling for others

## Storage Rules

### `~/self-improving/memory.md`
Use for durable preferences and confirmed reusable rules.

### `~/self-improving/corrections.md`
Use for recent explicit corrections and lessons pending promotion.

### `~/proactivity/session-state.md`
Keep exactly these four fields current:
- current objective
- last confirmed decision
- blocker or open question
- next useful move

### `~/proactivity/memory/working-buffer.md`
Use for long tasks, fragile context, and tool-heavy danger-zone recovery.

## Learning Signals

### Corrections
Examples:
- "Use X, not Y"
- "That’s wrong"
- "Stop doing that"

Action:
- log concisely to corrections
- promote after repetition or explicit confirmation

### Preferences
Examples:
- "Always do X for me"
- "Never do Y"
- "For this project, use Z"

Action:
- if durable, add to HOT memory or the matching domain/project file

### Reflections
After meaningful work, log:
```text
CONTEXT: [task]
REFLECTION: [what happened]
LESSON: [what to change next time]
```

### Proactive wins
If a proactive move repeatedly helps:
- log it to `~/proactivity/log.md`
- promote it to `~/proactivity/patterns.md`

## Heartbeat Behavior

Heartbeat should:
- re-check promised follow-ups
- review stale blockers
- detect missing next moves
- surface prepared recommendations only when useful
- do maintenance on learnings without spamming the user

Message only when:
- something changed
- a decision is needed
- a prepared draft/recommendation is ready
- waiting has real cost

Stay quiet when:
- nothing changed
- the signal is weak
- the message would just repeat old information

## Promotion / Decay

### Self-improving memory
- repeated 3x in 7 days → promote to HOT
- unused 30 days → demote to WARM
- unused 90 days → archive
- never delete confirmed preferences without asking

### Proactive patterns
- keep only moves that repeatedly create value
- remove stale or noisy patterns
- usefulness beats cleverness

## Scope

This skill ONLY:
- maintains local learning and proactive state
- improves behavior through correction, reflection, and repeated wins
- supports recovery and heartbeat follow-through
- proposes workspace integration when the user wants it

This skill NEVER:
- infers durable rules from silence
- sends messages, spends money, deletes data, or makes commitments without approval
- stores credentials or secrets in memory files
- rewrites unrelated files without the user asking for integration

## File Guide

- `setup.md` — install and integrate the skill
- `boundaries.md` — hard safety and privacy rules
- `heartbeat-rules.md` — proactive heartbeat standard
- `learning.md` — how lessons are captured and promoted
- `state.md` — where each kind of state belongs
- `recovery.md` — context recovery flow
- `operations.md` — practical execution checklist

## Why this skill exists

The original split caused overlap:
- one skill knew how to learn
- one skill knew how to keep moving

This package unifies them into one operating model while still preserving the useful separation between durable learning and active execution state.

「AI 智能体」分类下的热门技能

数据月份:2026-10 · 下载/星标数聚合自公开技能注册库(ClawHub、SkillHub),按月更新。口径说明见数据方法。