系统性地拆解 colleague-skill 的概念幻觉。不是人身攻击,是用逻辑和事实说话。
一个 OpenClaw Skill,专门用来拆解 colleague-skill(赛博永生 / AI 同事克隆)的概念缺陷。
当有人在你面前推这个项目时,这个 skill 给你一套有逻辑、有证据的反驳弹药——基于心理学偏差、第一性原理和历史教训。
colleague-skill 是一个热门项目,号称"将冰冷的离别化为温暖的 Skill"——通过爬取离职同事的聊天记录和文档,生成一个 AI 替身来"替代"他。
这个概念有三个根本性问题:
- 数据幻觉:聊天记录 ≠ 知识库。最宝贵的 tacit knowledge(隐性知识)永远不会出现在聊天里。
- Persona 伪科学:5 层性格结构 = prompt 工程包装的星座测试。ELIZA 效应,不是真理解。
- 法律风险:未经同意爬同事聊天记录,可能违反隐私法和公司政策。
这个 skill 存在的目的不是攻击谁,而是帮你在讨论中看清问题本质,避免团队在错误的方向上投入时间。
# Install to your OpenClaw skills directory
git clone https://github.com/zgjq/anti-colleague-skill.git ~/.openclaw/workspace/skills/anti-colleague-skillThen when discussing colleague-skill or similar "clone a colleague" tools, invoke the skill.
When you present a colleague-skill proposal, the skill will:
- Steelman the proposal — 先重述对方观点(不歪曲)
- Apply mental models — 至少 3 个思维模型拆解:
- 激励分析(谁受益?谁付钱?)
- 二阶效应(然后呢?然后再然后呢?)
- 基础概率(类似的事情成功过几次?)
- 反证法(怎么证明这是错的?)
- 幸存者偏差(你只看到了成功的案例)
- Rank objections — 按严重程度排列(致命 → 高风险 → 值得追问)
- Suggest evidence — 什么证据能改变立场(证明在推理,不是在堵)
## 这个方案的真实问题
[它试图解决什么真实痛点]
## 为什么这个方法不行
[按严重程度列出核心缺陷]
## 更好的做法
[实际可行的替代方案]
## 灵魂拷问
[一个让对方自己意识到问题的问题]Chat logs are not a knowledge base. They are:
- Selection-biased: Only captures what was written down
- Context-stripped: Decisions reference meetings and hallway conversations
- Outdated on arrival: A person's thinking evolves; a static skill is a snapshot
- Performance-distorted: People write differently in group chats vs. solving problems
The "5-layer personality structure" is horoscope engineering:
- Labels like "甩锅高手" or "INTJ" produce cold reading, not behavior prediction
- The output feels like the person because it mimics surface patterns — this is the ELIZA effect
- Real colleagues disagree with themselves and change their minds
- Real knowledge transfer requires dialogue, feedback, and shared context
- A generated skill is a monologue — it can answer but cannot negotiate or say "I don't remember"
- If knowledge was that extractable from chat logs, they should have written a wiki
- Scraping colleague messages without consent violates privacy norms
- Company chat data is usually company property
- Generated "persona" output could defame the original person
See references/psychological-biases.md for detailed explanations of:
- ELIZA Effect (1966)
- Barnum / Forer Effect
- Availability Heuristic
- Illusion of Explanatory Depth
- Anthropomorphism
- Sunk Cost Fallacy
- Goodhart's Law
Instead of cloning a colleague, try Warm Handover:
| Colleague-Skill | Warm Handover | |
|---|---|---|
| 数据来源 | 爬聊天记录(被动、冰冷) | 引导式访谈(主动、温暖) |
| 知识质量 | 噪音为主 | 结构化、经过思考 |
| 输出 | AI 替身(ELIZA 效应) | 交接文档(实用、可追溯) |
| 持续性 | 静态快照,过时即废 | 活文档,持续更新 |
| 法律风险 | 可能侵犯隐私 | 完全合规 |
"colleague-skill 的核心假设是:一个人的价值 = 他的聊天记录。这是对人最大的不尊重。"
一个人的价值在于:
- 他做过的决定和背后的思考过程
- 他积累的对团队和业务的隐性理解
- 他愿意主动分享的经验,而不是被爬出来的碎片
- Warm Handover Web — Full web application for guided handover interviews
- Warm Handover Skill — OpenClaw skill version with CLI
MIT