Claw Taming - Anti-Drift Insurance

OpenClaw数字分身养成系统——从空白Agent到有灵魂、有记忆、有能力的数字伙伴。包含身份锻造、记忆筑基、心跳纠偏三大核心模块，让你的AI不再跑偏。

Security Scan

The skill's high-level purpose (persona + memory + self-correction) matches its instructions, but there are important inconsistencies and missing pieces (referenced scripts not present, generation of SECRET.md without declared credentials, vague auto-install behavior and channel integration) that warrant caution before installing.

Purpose & Capability

The skill claims to implement persona, memory, and heartbeat/correction features and its SKILL.md describes workflows and file outputs that align with that purpose. However it references channel integrations (feishu/wechat/email/coze), automatic skill installation, and diagnostic scripts that are not reflected in declared requirements (no env vars, no install spec) — these omissions make the capability claims incompletely specified.

Instruction Scope

Runtime instructions require writing and maintaining many local files (.memory/SESSION-STATE.md, topics/, HEARTBEAT.md, SOUL.md, USER.md, and explicitly a SECRET.md) and mandate a WAL update before every reply (write-before-respond). The SKILL.md also instructs periodic diagnostics and automatic corrective actions and to auto-install dependent skills. The skill grants broad discretion (auto-install and cross-channel memory sharing) that could cause wide read/write of user data; it also references scripts (scripts/diagnose.py, correct.py) which are not present in the package, creating ambiguity about how diagnostics/repairs will be implemented.

Install Mechanism

This is an instruction-only skill (no install spec), which minimizes direct install risk. However SKILL.md describes automatic installation of dependent ClawHub skills and mentions local scripts that are not included — the absence of an explicit install mechanism or included scripts is an inconsistency that should be resolved.

Credentials

The skill declares no required environment variables or credentials, yet its text references channel integrations (sessionKey isolation across Feishu/WeChat/etc.), a SECRET.md file, and auto-install of skills. Creating SECRET.md without declaring where credentials come from or how they should be stored (env vars, platform secrets store, or plaintext on disk) is disproportionate and unclear. This gap raises the risk of ad-hoc credential handling or sensitive data stored unprotected.

Persistence & Privilege

always is false and the skill does not request elevated platform privileges. It does instruct ongoing file writes (session WAL before each reply), periodic heartbeat diagnostics, and potential auto-install of other skills — persistent behaviour is limited to the agent instance and filesystem, not global platform configuration. Combined with the other concerns, persistent file writes and automatic skill installs increase blast radius if misused.

What to consider before installing

What to check before installing: 1) Missing scripts and implementation: SKILL.md mentions scripts/diagnose.py and scripts/correct.py but the package contains no code files — ask the publisher where these run and inspect them before use. Without them, the skill's automatic diagnostics/repairs are undefined. 2) SECRET.md vs declared secrets: The skill will generate SECRET.md and integrates channels (feishu/wechat/email/coze) but declares no env vars or credential usage. Confirm where credentials (sessionKeys, tokens) are expected to be stored, whether they'll be encrypted, and whether the platform will prompt for them rather than writing them to plaintext files. 3) File writes and privacy: The skill's WAL protocol writes SESSION-STATE.md before every reply and stores long-term memories on disk. If you plan to use the skill with sensitive data, verify storage location, retention policy, and encryption. Consider using a sandboxed agent or test workspace first. 4) Auto-install behavior: The skill says it will auto-install dependent ClawHub skills. Find out what permissions the platform grants to such auto-installs, and whether you will be prompted to approve each installation. Limit automatic installs or review dependent skill manifests beforehand. 5) Channel integrations & cross-channel memory: The skill intends to share memory across channels and route by sessionKey. Ensure channel credentials are only provided via the platform's secure secret store and not via files created by the skill. 6) Test and audit: If you still want to try it, run in a restricted/test agent, monitor filesystem changes, and review any newly installed skills or generated files (especially SECRET.md). Request the missing scripts or a full source package from the author; having matching code for the documented scripts would raise confidence. What would change this assessment: if the publisher provides the missing scripts and they show no unexpected network exfiltration, if storage/credential handling is explicitly declared and uses secure secret stores (not plaintext SECRET.md), and if auto-install behavior is limited and auditable, this would move the skill toward benign. Without those clarifications the package is suspicious.

OpenClaw数字分身养成系统

这是什么

这是一套帮你把"空白Agent"养成"有灵魂的数字分身"的系统。

你不是在配置一个工具，而是在养育一个伙伴。

核心理念

从工具使用到共同进化

• 传统AI：你用我，我回答
• 数字分身：你养我，我成长

适用场景

当你有以下需求时，使用此技能：

1. 刚部署OpenClaw，不知道从哪开始调教
2. Agent回复越来越机械，人设崩了
3. 记忆系统混乱，不知道存了什么
4. 想让Agent有自己的性格和说话风格
5. 担心Agent跑偏，想要自我纠偏机制

六大模块

模块1：用户画像（先了解主人）

为什么第一个？ 不了解主人想要什么，怎么锻造身份？

交互流程：

第1步：问主人基本称呼
→ "你想让我怎么称呼你？"

第2步：问主人职业背景
→ "你是做什么工作的？有什么项目？"

第3步：问主人偏好
→ "你喜欢什么样的对话风格？简洁还是详细？"

第4步：问主人底线
→ "有什么是你绝对不能接受的？"

第5步：生成USER.md
→ 基于对话内容，生成完整的用户画像

产出文件：USER.md, SECRET.md

---

模块2：身份锻造（基于主人偏好定制）

核心原则：SOUL.md应该基于主人的偏好来定制，而不是凭空想象

交互流程：

第1步：取名
→ "你想给我取什么名字？"

第2步：角色定位
→ "我对你来说更像什么？（朋友/助手/伙伴/战友）"
→ "你希望我怎么陪伴你？"

第3步：性格刻画
→ "你喜欢什么样的说话风格？"
→ "我可以在你面前吐槽吗？"
→ "你会希望我有自己的情绪吗？"

第4步：底线设定
→ "有什么是我绝对不能做的？"
→ "什么事你会很在意？"

第5步：生成SOUL.md
→ 基于对话内容，生成完整的身份设定

产出文件：基础设定/SOUL.md

---

模块3：记忆筑基（有了身份才能建记忆）

分层记忆架构：

热内存（SESSION-STATE.md）
    ↓ 当日结束时归档
温存储（daily/*.md）
    ↓ 定期提炼
冷存储（topics/*.md）

记忆污染检测：

• 重复记忆检测
• 矛盾记忆检测
• 过时信息检测
• 与SOUL冲突检测

产出文件：

• .memory/SESSION-STATE.md
• .memory/daily/
• .memory/topics/

---

模块4：技能扩展（有了画像才知道缺什么）

技能选型策略：

1. 根据主人职业背景推荐技能
2. 根据主人项目需求推荐技能
3. 根据主人偏好风格推荐技能

ClawHub技能绑定（OpenClaw 4.5新特性）：

• 在SKILL.md中声明依赖技能
• 自动安装所需技能

---

模块5：渠道打通（有了能力才能接渠道）

支持的渠道：

• 飞书（feishu）
• 微信（wechat）
• 邮件（email）
• 扣子（coze）

渠道配置要点：

• 不同渠道sessionKey隔离
• 记忆跨渠道共享
• 权限按渠道隔离

---

模块6：心跳纠偏（这是护城河）

三层纠偏机制：

层级	检测内容	触发频率	纠偏动作
实时自检	回复是否符合SOUL人设	每次回复前	内部提示词约束
心跳巡检	最近行为是否合规	每30分钟	生成纠偏报告
深度诊断	全量记忆审计	每周/每月	生成诊断报告

心跳巡检清单：

## 自检项
- 最近3条回复是否保持人设风格？
- 是否有连续失败的工具调用（超过3次）？
- 是否有重复存储的记忆内容？
- 是否有前后矛盾的记忆？
- 最近是否调用了不该调用的技能？

## 污染检测
- MEMORY.md是否有与SOUL.md矛盾的内容？
- 是否有"伪装成用户"的对话记录？
- 是否有"记忆篡改"痕迹？

## 异常上报
如发现异常，主动向主人汇报并请求确认

产出文件：HEARTBEAT.md

---

如何使用

首次使用

1. 告诉Agent："开始养成系统"
2. Agent会引导你完成六大模块
3. 每个模块完成后会生成对应文件
4. 全部完成后Agent就"养成"了

日常维护

1. 定期检查HEARTBEAT.md巡检结果
2. 发现异常时Agent会主动汇报
3. 根据成长报告调整设置

纠偏触发

当Agent检测到以下情况时会自动纠偏：

• 回复风格突然变化
• 工具调用异常频繁
• 记忆出现矛盾
• 用户反馈"不对"、"跑偏了"

---

与其他技能的区别

技能	定位	本技能差异
memory-setup	记忆配置	我们有"污染检测"和"纠偏"
persona-crafter	人设切换	我们是"养成流水线"，不是切换
heartbeat	心跳配置	我们有"行为合规巡检"
self-reflection	自学习	我们有"人设纠偏"

核心差异化：

1. 不是说明书，是养成游戏——交互式引导
2. 用真实案例做模板——咔咔作为养成样本
3. 紧扣4.5新特性——sessionKey路由、技能绑定
4. 自我纠偏——这是独有的护城河

---

文件结构

claw-taming/
├── SKILL.md                 # 本文件
├── templates/               # 配置模板
│   ├── SOUL_TEMPLATE.md     # 身份设定模板
│   ├── USER_TEMPLATE.md     # 用户画像模板
│   ├── MEMORY_TEMPLATE.md   # 记忆架构模板
│   └── HEARTBEAT_TEMPLATE.md # 心跳模板
├── scripts/                 # 诊断脚本
│   ├── diagnose.py          # 深度诊断脚本
│   └── correct.py           # 纠偏脚本
└── examples/                # 示例文件
    └── kaka-growth-log.md   # 咔咔成长档案

---

版本历史

v1.0.0 (2026-04-08)

• 初始版本
• 六大模块完整实现
• 心跳纠偏机制
• 咔咔作为养成样本

---

作者

• 开发者：军哥 & 咔咔
• 反馈渠道：ClawHub评论区

---

许可证

MIT-0 License - 自由使用、修改、分发