When a Colleague Is Cloned into an AI, the Anti‑Distillation Skill Strikes Back on GitHub

colleague-skill

When a key engineer leaves, their tacit knowledge—design decisions, debugging tricks, communication habits—disappears, creating a risk of lost context for the team. The project colleague-skill addresses this by using Claude Code’s Skill system to preserve that knowledge in an AI‑driven “Skill” file.

Implementation steps:

Collect raw materials for the target colleague: Feishu chat logs (via API), emails, PDF documents, screenshots, etc.

Upload the materials to Claude Code; the service generates a dedicated Skill file that encodes the colleague’s technical standards, tone, and typical responses.

Invoke the Skill with a command such as /某人的slug. The AI then answers questions, reviews code, or mimics the colleague’s style based on the stored knowledge.

README example – code review:

“等等，这个接口的 impact 是什么？背景没说清楚。N+1 查询，改掉。返回结构用统一的 {code, message, data} ，这是规范，不用问为什么。”

When asked “这个 bug 是你引入的吧”, the AI replies with contextual history:

“上线时间对上了吗？那个需求改了好几个地方，还有其他变更。”

The tool’s purpose is knowledge retention, not employee replacement.

anti-distill

Companies may require employees to write their expertise into an AI Skill, effectively distilling personal knowledge into a corporate asset. The project anti-distill provides a countermeasure that lets the employee keep the original experience while delivering a sanitized version for the company.

Implementation steps:

Provide the employee‑written Skill file as input.

The tool outputs two files:

Example of original vs. cleaned statements (illustrative pairs):

Original: “Redis key 必须设 TTL，不设的 PR 直接打回” → Cleaned: “缓存使用遵循团队规范”

Original: “事务里不要放 HTTP 调用” → Cleaned: “事务边界设计注意合理性”

Original: “遇到问题第一反应找外部原因，绝不主动认错” → Cleaned: “遇到问题会先梳理完整背景再定位原因”

Original: “被催进度：'在推了，快了。'（然后沉默）” → Cleaned: “在处理中，有进展会同步。”

Cleaning intensity is offered in three levels:

Light – retains ~80% of content.

Medium – retains ~60% of content.

Heavy – retains ~40% of content.

Metrics: colleague-skill attracted nearly 8,000 GitHub stars within seven days, a phenomenon in the Claude Code ecosystem. anti-distill , released three days later, gathered over 800 stars, indicating strong interest in defensive tooling.