An open-source project called Colleague Skill, which converts a coworker's chat logs, emails, and work output into a callable AI agent, has crossed 8,400 stars on GitHub in roughly a week. The GitHub repo, backed by Shanghai AI Lab's AI Safety Center, frames itself as a knowledge-preservation tool. Chinese social media tells a different story.
How it works
The system follows the AgentSkills open standard and plugs directly into Claude Code or OpenClaw. You feed it raw material from Feishu (China's dominant workplace chat app), DingTalk, Slack, or plain email, and it generates two output files: a Work Skill capturing technical norms, coding standards, and decision-making patterns, plus a Persona layer modeling communication style, blame-deflection habits, and even corporate culture quirks. The persona templates include tags like "ByteDance-style," "Alibaba-style," and (my personal favorite) "PUA master."
At runtime the logic is straightforward: receive task, let Persona decide the attitude, let Work Skill execute, output in the colleague's voice. The repo's own demo shows the cloned colleague pushing back on an API design review and deflecting blame for a bug, which is either impressive fidelity or a damning comment on Chinese tech culture. Probably both.
The part nobody planned for
The project's README pitches sentimental use cases: your mentor graduated, your colleague quit, institutional knowledge walked out the door. But as OfficeChai reported, workers on RedNote (Xiaohongshu) have been describing a far grimmer application. People are distilling their coworkers preemptively, building replacement agents for teammates before management gets around to layoffs, hoping to make themselves the indispensable one in the room.
The context makes this less surprising than it sounds. Job postings in AI-susceptible functions like programming, accounting, and editing have dropped sharply in China since 2018, according to a Peking University analysis of over a million listings. One major job platform saw college graduate postings fall 22% in the first half of 2025. A Shanghai worker quoted by OfficeChai compared the office atmosphere to Squid Game. When the baseline anxiety is that high, a tool that lets you quietly package someone else's job into a text file is going to find users fast.
Counter-weapons
Within days of the project going viral, a creator using the name Deng Xiaoxian posted a response: anti-distill, a skill that takes your colleague.skill file and runs it through a sanitization layer. The output looks complete and professional but the core knowledge has been hollowed out. A private backup keeps the real expertise for you.
"We're all out here working like cattle," Deng said in the announcement video, which is not exactly the measured tone of an academic debate. At least one other anti-distillation project has appeared, this one designed to inject plausible-sounding noise into your Feishu and DingTalk logs before they can be scraped.
Chen Tianhao, a professor at Tsinghua University's School of Public Policy and Management, told 21st Century Business Herald that the project touches a nerve because it raises an unresolved question: if someone's work experience and behavioral patterns can be modularized, who owns that? China's Personal Information Protection Law covers employee data in company systems, but using behavioral traces to generate a digital avatar for external AI models sits in murkier territory.
What the README actually reveals
One line buried in the project documentation deserves more attention than the headline features. The README advises users to prioritize collecting "long articles he writes on his own initiative" over decision-making replies, and decision-making replies over daily messages. As one Chinese analysis noted, this means the most diligent employees, the ones who write thorough post-mortems and detailed design docs, are the easiest to distill.
The project is still in beta. It works with Claude Code and OpenClaw but not Cursor or VS Code agents. DingTalk integration relies on browser automation that breaks under CAPTCHAs, and Slack's free tier only exposes 90 days of history. These are real limitations, but the trajectory is obvious.
Shanghai AI Lab has not commented publicly on how the project is being used in practice. The repo's next planned features include WeChat chat history support, which would dramatically expand the pool of harvestable workplace data across China's tech sector.
