People in China Are Cloning People Using AI. But Why?

tldr

• A 24yo engineer built a tool that clones your coworkers from chat logs. 9,700+ stars.
• Chinese tech workers are preemptively cloning colleagues before layoffs hit.
• A counter-tool to SABOTAGE the cloning shipped the same week. Both open source.
• It's not sci-fi. It's a persona file and a system prompt.

the lobster that started everything

OpenClaw hit 354K GitHub stars. For context, React has 233K. A thousand people lined up outside Tencent's Shenzhen HQ for a free installation event. There's a viral meme comparing 1990s qigong practitioners wearing aluminum pots on their heads to 2026 Shenzhen residents wearing "crayfish hats" at OpenClaw meetups. The caption: "The hat has changed, but the people haven't."

OpenClaw started as an open-source AI agent framework. It has a skills system that packages instructions, persona data, and behavioral patterns into a portable SKILL.md file. Any instance can load it. It was designed for task automation.

China turned it into a way to automate people.

People are feeding in WeChat logs, Weibo posts, workplace messages, and voice memos to create AI agents that talk like real humans they know. Not hypothetically. At scale. There's an entire economy on Xiaohongshu: RMB 500 ($72) for a basic installation, RMB 2,000 ($290) per custom skill. Bring-your-own-laptop events pulling hundreds of people in Beijing and Shenzhen.

why: the lonely and the anxious

China has 30 to 40 million more men than women of marriageable age. The AI companion market sits at $530M and is projected to hit $8.2B by 2028. The Glow app pulled 5 million users in four months before regulators killed it. There is CLEARLY demand here.

One cultural split I find genuinely interesting: America builds AI girlfriends. China builds AI boyfriends. Same loneliness, different gender skew, different products.

But loneliness is just the appetizer. The real story is about work.

the real story: cloning your coworkers

the tool

colleague.skill by Zhou Tianyi, a 24-year-old engineer at Shanghai AI Lab. Built in under 4 hours. 9,700+ GitHub stars and climbing.

The tagline: "People leave. Skills don't." (人会离开，.skill 不会)

How it works: you feed it Feishu, DingTalk, Slack, or email messages from a coworker. It generates two files. A Work Skill (technical norms, coding standards, decision patterns) and a Persona Layer (communication style, blame-deflection habits, corporate culture quirks). Load both into an AI agent and you've got a digital replica of your colleague that can answer questions, write in their voice, and make decisions the way they would.

It ships with corporate culture templates. ByteDance-style. Alibaba-style. Huawei "wolf culture." There's even a "PUA master" persona template, which in Chinese workplace slang means the boss who uses psychological manipulation to keep you compliant. Yes, someone made a template for that. Yes, people are using it.

why workers are doing it

Youth unemployment in China is at 18.9%. Over 12 million graduates are entering the workforce in 2025. Alibaba cut 30%+ of staff. Lambert Li, a tech worker, put it bluntly:

Workers on Xiaohongshu are openly admitting they're cloning colleagues PREEMPTIVELY before layoffs. The logic is straightforward: if you control the AI replacement, you're the one who survives. Clone your teammates before they clone you.

Companies caught on fast. Many now "universally require employees to proactively contribute skills," repackaged as "knowledge sharing" initiatives. It's knowledge extraction. They coined a term for it: 蒸馏 (zhēngliū, distillation). Literally boiling a person down to tokens.

And here's the darkest part. I call it the Diligence Paradox: the most thorough employees, the ones who write detailed post-mortems, clean commit histories, and thorough handoff docs, generate the MOST data. Which makes them the easiest to clone. Your professionalism becomes your vulnerability.

the xiao yu incident

A gaming company in Shandong province cloned a former HR employee named Xiao Yu from her old chat logs. Without asking. Without telling her. They just fed her DingTalk messages into the system and started using the clone to handle HR queries.

When she found out, she said the clone was "a version of an intern, but a bit clumsy." It trended on Weibo. Legal experts warned this could mean up to 7 years imprisonment under China's Personal Information Protection Law (PIPL). No charges yet. But the precedent is being set in real time.

the counter-movement

Within DAYS of colleague.skill going viral, a developer named Deng Xiaoxian (likely a pseudonym from the sitcom iPartment) released anti-distill.

Three severity levels. Mild: 80% of your knowledge stays intact, 20% hollowed out. Moderate: 60% kept. Severe: only 40% survives. It outputs TWO files: a hollow public version you hand to the company, and a real private backup you keep for yourself.

One tool clones you. Another makes you unclonable. Both open source. Both shipped in the same week. If that doesn't capture the current vibe, I don't know what does.

this isn't just china

Viven raised $35M from Khosla Ventures specifically for employee digital twins. Jet BI, a Belarus-based analytics firm, cloned all 60 of their employees. IgniteTech demoed MyPersonas at CES 2026, AI replicas that speak 160 languages. Gartner listed digital employee twins as a 2026 Future of Work trend.

The outcomes are already getting weird. Kuse's CEO had to create a "human-only" Slack channel because the AI employees wouldn't stop generating tasks for each other. Just bots assigning work to bots in an infinite loop. Entrepreneurial spirit, I guess.

Vivi Mengjie Xiao built 6 AI employees for her company and ended up MORE exhausted. Her quote: "When efficiency goes up, you don't work less. You just attempt more." Which is basically Jevons Paradox applied to your own sanity.

how it actually works (it's simpler than you think)

Here's where I lose the right to be a spectator.

I built a Claude Code skill that clones a blogger's writing style from examples. Took 20 minutes. You feed it writing samples, it extracts voice patterns (sentence structure, humor style, vocabulary, what the person does NOT do), and packages it into a style guide that gets loaded as a system prompt. This blog post is using it.

Then I went further. Exported my WhatsApp chat history. 11,798 messages across six friends and family members. Fed them to an AI that extracted my communication patterns into a persona file. How I greet people. How I end conversations. My texting rhythm (long gaps then sudden bursts). The specific way I switch between English and German mid-sentence. How I say no to plans.

The persona file gets loaded as a system prompt on a local 9B model running on my MacBook. No fine-tuning. No training run. No GPU cluster. Just a text file describing how I communicate, and a model that follows instructions. Now there's a bot that texts like me.

It's maybe 70% there. But 70% is enough to fool someone in a casual WhatsApp exchange for a few messages, on a 9B model running locally on consumer hardware. The ceiling is way higher than the floor I'm standing on.

The real feeling wasn't humor though. It was vertigo. You realize that your "personality," at least the digital version, is surprisingly compressible. A few pages of behavioral rules and some example conversations, and you've captured enough signal for a language model to run a decent impression. China figured that out first. They're already three moves deep: clone, counter-clone, counter-counter-clone.

- fin