How to Build a Colleague‑Mimicking AI Agent with Claude Code

The author presents colleague‑skill , an open‑source Python tool (MIT‑licensed) that harvests raw materials such as Feishu, DingTalk, Slack, email, PDF screenshots, and Markdown to create reusable AI Skills capable of mimicking a specific colleague in Claude Code.

The repository has attracted over 8.8k stars and 700+ forks, indicating strong community interest. It is maintained by the Shanghai AI Lab and AI Safety Center, and continuously updated rather than being a one‑off marketing repo.

Each colleague Skill is divided into two parts: Work (formal process knowledge) and Persona (a five‑layer personality model). Incremental files can be merged, and corrections can be applied instantly with version rollback. The system supports multiple data sources and includes a tag library that captures the cultural nuances of major Chinese tech firms.

Technical core: a Python multi‑channel collector parses the raw inputs into text, runs a prompt pipeline that separates the content into Work and Persona. When a task arrives, the Persona sets the tone, then Work executes the action, producing responses that sound like the actual colleague.

Effect Examples

Input: 字节 2-1 后端工程师，INTJ，甩锅高手，字节范

Scenario 1: Code Review

用户      ❯ 帮我看一下这个接口设计

同事.skill ❯ 等等，这个接口的 impact 是什么？背景没说清楚。
          （看完后）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。

Scenario 2: Blame Shifting

用户      ❯ 这个 bug 是你引入的吧

同事.skill ❯ 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。

To use the skill, clone it into the Claude skills directory (e.g., .claude/skills/create-colleague). Optional dependencies can be installed, and full Feishu/DingTalk automation requires app credentials; otherwise, you can manually export files. pip3 install -r requirements.txt In Claude Code, run /create-colleague, fill in name, level, tags, and select data sources. After creation, invoke the skill via /{slug} for listing, rollback, or deletion.

The project is in beta with around 60 open issues, runs on Python 3.9+, and includes a PDF technical report that details the two‑layer architecture for deeper exploration.

Pros : captures undocumented knowledge, enables a “cyber colleague” for code review or other interactions, and offers a fun way to study skill engineering.

Cons : the skill’s quality depends on the source material; users bear all privacy and compliance risks and must avoid uploading sensitive chat logs to public servers.

Repository: https://github.com/titanwings/colleague-skill

Technical report PDF: https://github.com/titanwings/colleague-skill/blob/main/colleague_skill.pdf