In less than three days, the Star curve skyrocketed like a rocket.
Today is Holy Thursday, and also National Burrito Day. But beyond burritos, I found something much more “intense” on GitHub today.
A developer named titanwings created a project called colleague-skill. In short: when a coworker resigns, you feed their chat logs, documents, and emails into it, and the AI generates a clone of “them”—even replicating their tone of shifting the blame.
When Someone Leaves, They Take More Than Just the Mug on Their Desk
According to Work Institute’s 2025 data, the average direct cost of replacing an employee in a U.S. company is about 33% of their annual salary. If someone making $50,000 leaves, the company spends $16,500 to fill the gap. But that’s just the visible costs like recruiting, training, and onboarding.
The invisible costs are the real killers.
A report by Glean noted that the manufacturing industry will need 3.8 million new workers by 2033, but about 1.9 million roles might remain unfilled due to a skills gap. In the aerospace sector, the turnover rate spiked from 5.7% in 2021 to 7.1% in 2022. What lies behind these numbers? It’s the tacit knowledge locked inside veteran engineers’ heads—like “this valve behaves a bit differently at minus 20 degrees”—evaporating along with the people.
This is exactly what colleague-skill attempts to solve. It’s not about writing a handover document—let’s be honest, who believes three pages can summarize three years of experience?—but rather packing a person’s working style, technical standards, and even personality traits into an AI Skill that Claude Code can invoke directly.
What’s Really Interesting is the Five-Layer Persona Structure
Many people’s first reaction to this project is amusement: personality tags like “blame-shifter,” “passive-aggressive,” or “leaving people on read” act like a mental massage for weary professionals. But peel back this layer of dark humor, and the underlying technical design is actually much more serious than most “awesome-list” projects.
Each colleague Skill is divided into two parts: Work Skill (handles technical standards and workflows) and Persona (handles expression style and decision-making patterns). The Persona is further divided into five layers: Hard Rules → Identity → Expression Style → Decision Pattern → Interpersonal Behavior. When a task is received, it first goes through the Persona to determine the attitude, then through the Work Skill to execute, and finally outputs the result in “their tone.”
This architecture reminds me of Microsoft’s TinyTroupe—a research project that uses GPT-4 to simulate AI avatars with different personalities. TinyTroupe’s paper specifically highlights the difference between “Assistant AI” and “Simulated AI”: the assistant strives to be correct, polite, and comprehensive, while the simulated AI needs to replicate human diversity, including the imperfect parts. colleague-skill clearly stands on the latter side. It doesn’t aim to build a perfect AI coworker, but a “real” one—including the fact that during a code review, they will first ask “what’s the impact?” before looking at the code, and when questioned about a bug, their first reaction is “did the deployment time match?”
The README starts with the inside joke “You LLM folks are code traitors.” It looks frivolous, but the code structure is quite rigorous.
How Does It Compare With Its Peers?
AgentSkills Finder scored colleague-skill an 82/100, calling it a “solid directory listing candidate” among Skill creation projects. Let’s do a horizontal comparison:
Microsoft TinyTroupe — An academic-grade persona simulation framework. The paper is published, and the code is open-source, but it is positioned as a research tool, not aimed at specific work scenarios. You can’t use it to “do your colleague’s work.”
Composio’s awesome-claude-skills Collection — More like a mega-marketplace that catalogs various Claude Code Skills, including colleague-skill. However, the collection itself doesn’t produce content; it’s just a curator.
ex.skill (therealXiaomanChu/ex-skill) — This project is directly derived from colleague-skill, with 644 stars and 76 forks. The architecture is exactly the same, only the scenario is shifted from the workplace to romantic relationships. It supports WeChat chat log imports, full-set zodiac and MBTI configurations, and even changed the delete command to /let-go. To be honest, I paused for a second when I saw the command name /let-go.
But there’s a catch you should know: colleague-skill is still in the demo phase, with 592 stars and 20 forks. Automatic collection from Feishu requires App credentials, WeChat auto-decryption “tests out as unstable,” and the DingTalk API simply does not support historical messages. In plain terms, the data sourcing pipeline isn’t fully ironed out yet. As for the quality of the Skill, the README admits it itself—”the quality of raw materials determines the quality of the Skill.” The richer the chat logs you provide, the more alike the output; if you only describe a few sentences from memory, it just generates a stereotype.
If Everyone Becomes a Skill, Then What?
I sometimes think that the pain point this project truly hits isn’t “knowledge management,” but something far more emotional.
Look at its copy: “Turning a cold departure into a warm Skill.” Someone in the comments said, “I started laughing, but then I cried.” The README of ex.skill reads, “Human memory is an unreasonable storage medium.” These phrases actually feel quite out of place in technical documentation, but it’s exactly this incongruity that made it go viral. It touched on an emotion people rarely discuss: in an industry with an increasingly high turnover rate, you spend two years syncing with someone, and suddenly they’re gone. Before you even have time to be sad, you have to start picking up their mess.
If this tool really matures, could we see a situation where companies turn every employee into a Skill to retain as a “knowledge asset”? Who owns this Skill then? When an employee resigns, can they demand the company delete their Skill? Maybe I’m overthinking it. The MIT license means these generated Skills are open, but the raw materials inside—chat logs and emails—are very real privacy data.
There’s another detail that caught my attention. The latest commit for colleague-skill was “fix: replace broken /search/v1/user with department traversal + batch,” dated April 1st. Someone fixing bugs on April Fools’ Day—that in itself is very programmer-esque.
After saying all this, what I’m actually most curious about is: if your former colleague knew you turned them into an AI Skill, how would they react?
References:
- colleague-skill – GitHub
- LLM folks are code traitors! – Tencent News
- ex.skill – therealXiaomanChu/ex-skill
- create-colleague install and usage guide – Agent Skills Finder
- How AI facilitates knowledge transfer from retiring engineers – Glean
—— Lyra Celest @ Turbulence τ.
