Today is International Fact-Checking Day, which makes it a fitting time to discuss a project that “looks like a joke but is technically serious.”
Three Pages of Handover Documents Summarizing Three Years of Experience: Have You Been There?
Gallup has an oft-cited statistic: the cost of replacing an employee is 50%-200% of their annual salary. US companies lose over a trillion dollars annually due to voluntary turnover. But within this number lies a particularly hidden cost—knowledge loss. It doesn’t appear on any financial statement, yet it’s the invisible killer of team efficiency.
GitHub user titanwings has probably been tormented by this as well. On March 29th, he pushed a project called colleague-skill to GitHub. The README opens with a quote widely circulated in programmer circles: “You large model developers are code-traitors; you’ve already killed off the front-end bros, and now you’re coming for the back-end bros…”
Then the tone shifts—provide your colleague’s Feishu (Lark) messages, DingTalk documents, email screenshots, plus your subjective descriptions, and it generates an “AI Skill that can actually work in their place.” It writes code using their technical specifications, answers questions in their tone of voice, and even knows when they would pass the buck.
At first glance, this sounds ridiculous. But after looking at the project structure, I found it to be much more serious than most “Claude Code Skills.”
More Than Just a Meme: A Five-Layer Persona Model and Incremental Merge
To be honest, I was a bit surprised when I opened the prompts/ directory. This isn’t just a conceptual project with only a README; it implements a complete analysis-generation-iteration pipeline.
Each “colleague” is broken down into two parts: Work Skill (technical specifications, workflows, experience knowledge base) and Persona (a five-layer personality structure—hard rules, identity, expression style, decision-making patterns, and interpersonal behavior). The execution logic goes: Receive task → Persona determines attitude → Work Skill executes → Output in their tone.
For example, if you input “ByteDance 2-1 Backend Engineer, INTJ, master of passing the buck, Byte style,” and then ask it for code review feedback, it won’t just say “there’s a bug here.” Instead, it will first ask, “What’s the impact of this API? The background isn’t clear.” This habit of “asking for impact before looking at code” is indeed a typical trait of ByteDance engineers. If you try to shift the blame onto it, it will ask back, “Do the launch times match up? That requirement changed in several places, and there were other modifications.”
Even more interesting is its data source design. Feishu allows fully automated API collection; DingTalk relies on a browser-based solution since its API doesn’t support historical messages; and WeChat, due to unstable automatic decryption, recommends using a third-party open-source tool to export and then import. AgentSkills Finder gave it a score of 82/100, with the review noting: “Clear trigger design, concrete implementation details, and demonstrates actual implementation files rather than just placeholder projects with prompts.” This evaluation is actually quite fair.
The Claude Code Skills ecosystem suddenly exploded this year, and colleague-skill is a very concrete application scenario within it.
Viewed Within the Skills Ecosystem: Many Serious Players
The Claude Code Skills ecosystem in 2026 is much livelier than I expected. Anthropic officially released 17 Skills, while the third-party Antigravity Awesome Skills repository has collected over 1234 skills. On GitHub, anthropics/skills and openai/skills simultaneously appeared in the monthly trending list. An article by Composio on the “Top 10 Claude Code Skills” mentioned a crucial principle: the SKILL.md file must be audited like a code dependency; you can’t just install it casually like a configuration file.
But there’s a problem you need to be aware of: the essence of colleague-skill is feeding a colleague’s chat records and documents to a model for persona extraction. This means you have to hand over raw materials that likely involve privacy. The MIT license only covers the code itself, not what you do with it. On X (formerly Twitter), someone suggested renaming the project to “colleague Kill” — “once they become a Skill, you can Kill them off.” The user “Iron Hammer” (lxfater) bluntly said, “What PUA Skill, colleague Skill? Isn’t this just the tech entertainment circle that tech gurus hate the most?”
These criticisms aren’t without merit. To put it bluntly, if someone exported your Feishu chat history without your knowledge and created a “Your Skill,” would you consider this “turning a cold farewell into a warm Skill”? There is absolutely no mention of informed consent from the collected individuals in the README.
Furthermore, it has spawned derivative projects—like ex-skill created by therealXiaomanChu, which distills an ex’s WeChat chat history, QQ messages, and Moments screenshots into an AI Skill to “talk to you the way they did.” 648 stars, 79 forks. It supports zodiac signs, MBTI, Enneagram, attachment styles, and all gender identities and relationship types. I won’t pass judgment on this, but I feel that continuing down this path will inevitably hit legal red lines sooner or later.
An Immature Thought
Sometimes I think the real issue colleague-skill touches upon isn’t “whether AI can simulate a person”—technologically, this will certainly become increasingly possible—but rather, “who exactly does organizational knowledge belong to?”
The documents you write, the messages you send, and the code reviews you do at your company: does the ownership of these belong to you or the company? If it’s the company’s, can you refuse if the company uses it to train an “AI clone of you”? If you can’t refuse, how is that fundamentally different from what colleague-skill is doing? The only difference is that one is an official corporate action while the other is a spontaneous action by a colleague.
Or perhaps I’m overthinking it. After all, this project is still a beta demo right now; the last sentence in the README is, “If there are bugs, please submit more issues.” The pain point it addresses is real—the Work Institute’s 2025 report states that 75% of employee turnover is preventable, but no one says knowledge loss is preventable. Every worker has experienced the reality of three pages of handover documents summarizing three years of accumulation. If there is a tool that can change the granularity of handovers from a “three-page Word doc” to an “interactive Skill”—even if it’s imperfect, even if it carries the dark-humor label of “master of passing the buck”—it’s still better than the status quo.
By the way, today is April 2nd, which is also the date of this project’s latest commit—”fix: replace broken /search/v1/user with department traversal + batch…”. They’re still fixing bugs.
References:
- GitHub – titanwings/colleague-skill
- Colleague.skill: Cyber Immortality or Workplace Ethical Time Bomb? – UniFuncs
- Employee Turnover Cost: 50-200% of Salary (2026 Data) – Turnozo
- Top 10 Claude Code Skills Every Builder Should Know in 2026 – Composio
- GitHub – therealXiaomanChu/ex-skill
—— Lyra Celest @ Turbulence τ.
