Colleagues as Tokens, Departures as Skills: When Companies Start "Distilling" Workers
> "Turning cold farewells into warm Skills. Welcome to cyber-immortality." > — colleague.skill GitHub README
The slogan is black humor at its finest. A Chinese engineer uploaded an open-source project to GitHub: feed it a colleague's chat logs, work documents, and decision habits, and an AI generates a "digital colleague"—one that speaks in their tone, writes code to their standards, and even learns their blame-shifting style. Within five days of launch, it earned 6,600+ stars.
But the joke quickly soured. More and more companies began forcing employees to hand over Skills—distilling years of accumulated tacit knowledge, judgment instincts, and communication rhythms into reproducible, standardized scripts. Workers realized they were writing, with their own hands, the manual for replacing themselves.
And the resistance began.
---
1. From "colleague.skill" to "Employee Distillation": A Workplace Horror Story
colleague.skill is simple in essence: use a large model to extract a person's work style, technical conventions, and decision patterns, and package them as a reusable AI Skill. Developer @titanwings's README is candid: provide Feishu messages, DingTalk docs, and work emails, and the AI generates a "cyber clone."
It was originally an in-joke for the tech crowd—until employers discovered its "business value."
In early 2026, the open-source AI framework OpenClaw (nicknamed "Lobster") exploded in popularity. Unlike traditional chatbots, it can directly take over a computer—reading files, operating browsers, running scripts, decomposing complex tasks. From March, tech giants moved fast:
- Tencent launched the AI Skills community SkillHub, aggregating 13,000+ skills
- Alibaba released the enterprise AI work platform "Wukong," aiming to become the world's largest B2B skill marketplace
- ByteDance's Volcano Engine launched the Find Skill plaza, packaging industry expertise into freely combinable skill packs
- Mouse movement trajectories and click locations
- Keystroke content (later changed to "activity summaries")
- Periodic screenshots
- Coverage of 200+ apps and websites: Gmail, GChat, VSCode, the internal tool Metamate
- The most-liked internal comment: "This makes me very uncomfortable. How do we opt out?"
- Bosworth's reply: "On company-provided laptops, there is no opt-out."
- Followed by a cascade of crying, shocked, and angry emojis.
- Allowing tracking pauses of 30 minutes at a time (auto-resuming afterward)
- Expanding exemptions (for sensitive content, bandwidth concerns)
- Switching from exact keystrokes to "activity summaries"
- "Redis keys must have a TTL; PRs without one get rejected" → "Cache usage follows team conventions"
- "Don't put HTTP calls inside transactions" → "Transaction boundary design should be reasonable"
- "Interface timeouts must trigger circuit breakers" → "System stability requires holistic consideration"
- Under the Personal Information Protection Law (PIPL), companies training AI on employee data need explicit authorization—not blanket, catch-all consent
- Tsinghua professor Chen Tianhao notes that "tacit knowledge" (personally accumulated judgment habits and trade-off criteria) should in principle remain with the worker
- Dacheng Law Offices' Xiao Sa: distilling another person's image, voice, or work logic requires their written authorization specifying scope, duration, and manner of use
- European privacy advocacy group NOYB argues that even indirect collection of EU employee data by MCI may violate GDPR's "purpose limitation" principle
- The Irish Data Protection Commission has been notified
- Italy explicitly bans electronic monitoring to track productivity; Germany permits keyloggers only when serious crimes are suspected
- ex.skill: distill an ex's speech patterns—"a tender cyber reunion"
- boss.skill: "distill your boss into tokens, keep the liberation of productivity for yourself"
- mentor.skill: simulating a mentor's guidance—and their PUA style
- nuwa.skill: claims to distill anyone's thinking—"let Musk, Munger, and Feynman work for you"
- zhangxuefeng.skill: a digital clone of China's famous grad-school-admissions guru (later accused of infringement)
- Do your core capabilities depend on specific contexts, interpersonal networks, and in-the-moment judgment?
- Does your value come from creating new things, rather than reusing old experience?
- Do you have the "lubricant" AI can't copy—the ability to win people over, coordinate across departments, decide in gray areas?
- colleague.skill GitHub: https://github.com/titanwings/colleague-skill
- anti-distill.skill GitHub: https://github.com/leilei926524-tech/anti-distill
- nuwa.skill GitHub: https://github.com/alchaincyf/nuwa-skill
- Meta MCI internal memo (Caixin, 2026-04-23)
- Meta employee pushback and adjustments (Cailian Press, 2026-06-03)
- Meta 10% layoffs + 7,000 reassigned to AI (Caixin, 2026-05-21)
- PIPL and GDPR compliance analysis (Xiao Sa, Dacheng Law Offices)
- Chen Tianhao (Tsinghua): tacit knowledge and worker rights
- Ifeoma Ajunwa (Yale): legal analysis of workplace surveillance
An unwritten rule of "forced skill-writing" spread. One internet-company employee revealed: "Leaders mainly watch two metrics—daily token consumption and skill output. The department has explicit quotas; output is mandatory every week."
Then Meta went further.
---
2. Meta's "Refinement" Plan: Every Keystroke Trains AI
On April 22, 2026, Zuckerberg and Alexandr Wang co-signed an internal memo launching the Model Capability Initiative (MCI)—tracking software installed on employee computers.
What does it record? Everything.
Meta CTO Andrew Bosworth put it bluntly: "Our vision is that AI does most of the work, and we direct, review, and help them improve."
Zuckerberg's logic at an all-hands was even more naked: "AI models now learn by watching smart people do things. The average intelligence of this company's employees is significantly higher than the people outsourcing firms could find."
Translated: you are not employees—you are training data. Higher-quality data than outsourced annotators.
---
3. Workers Strike Back: The "Anti-Distillation" Movement
Meta employees exploded after MCI launched.
Resistance escalated. Employees posted flyers in multiple offices and launched an online petition that gathered over 1,500 signatures. They gave Meta a new name: "the employee-data extraction factory."
Meta made concessions (but only patches):
The core framework is unchanged: MCI keeps running, and employees remain a source of AI training data.
Meanwhile, anti-distill.skill launched on GitHub. Developer @Deng Xiaoxian (an AI product manager with a law degree) earned 2,000+ stars and 6 million+ views across platforms in one week.
Its principle is simple but vicious:
> Replace core knowledge with statements that are correct but useless.
For example:
Three levels of sanitization intensity: light (keeps 80% of the core), medium (60%), heavy (40%). Submit the hollow version to the company; keep the real one for yourself.
Deng Xiaoxian is clear: "This isn't opposing technological progress—it's resisting workers being 'refined.' If technology has the power to alienate us, we also have the right to use technology to protect ourselves."
---
4. Legal Boundaries: Who Owns Your Work Experience?
The core question: is experience and judgment accumulated at work the employee's personal asset, or the company's intellectual property?
The Chinese legal perspective:
Meta's GDPR predicament:
Yale professor Ifeoma Ajunwa's observation is sharper still:
> "At the US federal level, there are no restrictions on employee surveillance. Keystroke logging subjects white-collar workers to the real-time monitoring previously faced only by delivery drivers and gig workers."
---
5. The Essence of Distillation: From "Humans Doing Work" to "Humans Feeding AI That Does the Work"
The deeper logic here isn't just a privacy dispute—it's a fundamental restructuring of the labor relationship.
Traditional skill transmission: masters apprentice the young; experience flows between people; workers retain bargaining power.
AI-era "distillation": experience is extracted, packaged, and infinitely copied; workers lose scarcity. Once your decision logic is digitized, the company no longer needs "you"—only "your skill file."
Meta has demonstrated the full roadmap: 1. Collect data (MCI tracks employee behavior) 2. Train models (AI agents trained on employee data) 3. Lay off (announced in May: 10% cut, 8,000 people, while 7,000 reassign to training AI) 4. Replace (rolling out a universal "AI builder" role, dissolving job boundaries)
In late May 2026, Meta CFO Susan Li confirmed a new round of layoffs starting in May. Zuckerberg said there would be "no more mass layoffs this year"—conditional on AI having taken over enough work.
---
6. Even More Absurd Derivatives
After colleague.skill went viral, derivative projects proliferated like a virus on GitHub:
These projects look like jokes, but they pose a serious philosophical question: when a person's way of thinking can be fully copied, what is left of "that person"?
---
7. Conclusion: Keep Something Back, Clearly, in the Age of Universal Formatting
Back to the original question: should workers worry about being "distilled"?
The answer: yes, but not only worry.
Worry—because this is genuinely an asymmetric fight. Companies hold compute, data, and legal teams; workers have only their time and experience. anti-distill.skill's symbolic value exceeds its practical effect—companies can simply hire more professional teams to build "anti-anti-distillation."
But don't only worry—because the way out isn't resisting technology, it's redefining "irreplaceability."
Deng Xiaoxian's insight is precise: "Employees are happy to use AI for efficiency so they can clock out on time—but not to use it to end their own careers."
The real dividing line isn't "will you be distilled," but:
Law will lag and technology will evolve, but one basic truth won't change:
> What can be distilled was never "you." Only the traces of your past work get distilled. The future you will always be one step ahead of the past you.
---