colleague-skill: AI clones of colleagues for Claude Code

colleague-skill: Creating Digital Clones of Colleagues for Claude Code and OpenClaw

The colleague-skill project on GitHub, under the MIT license, implements the creation of digital clones of specific employees based on Claude Code and OpenClaw. The system analyzes correspondence from Feishu, DingTalk, documents, emails, and screenshots, supplemented with a profile description like "ByteDance, grade 2-1, backend, INTJ, master of shifting blame". The result is an agent that mimics the original's coding style, communication, and behavioral patterns.

The clone generates code according to the employee's personal technical preferences, responds in their tone, and applies characteristic strategies for dodging responsibility. The author highlights a provocative manifesto against replacing developers with large models.

Clone Architecture: Work Skill and Persona

The clone is divided into two components:

• Work Skill: stores the technical profile — preferred libraries, coding patterns, and compliance with guidelines.
• Persona: a character model in five layers:

1. Strict behavioral rules.

2. Identity and self-description.

3. Speech and communication style.

4. Decision-making method.

5. Interaction with people.

When receiving a task, Persona determines the emotional context, Work Skill handles the technical part, and the response is formatted in the original's style. The system supports incremental learning: adding files, manual adjustments with versioning and rollback.

Profile tags are adapted to Chinese corporate cultures: "ByteDance vibe", "Alibaba vibe", grades from ByteDance/Tencent. Behavioral tags include "master of shifting blame", "passive-aggressive", "plays office politics".

Clone Behavior Examples

In the demo of the "ByteDance 2-1 backend INTJ, master of shifting blame" clone, when accused of a bug, the agent responds:

Do the release dates line up? That task went through a ton of changes, and there were other edits too.

This illustrates the integration of Persona into responses. The project is positioned as a demo with an open invitation for bug reports.

The skill follows the AgentSkills standard, simplifying integration into Claude Code and OpenClaw ecosystems. The neologism "码奸" (code traitor) underscores criticism of AI replacement of specialists.

Integration and Customization

For deployment, collect data on the colleague and provide a profile in description format. The system automatically vectorizes the context and fine-tunes behavior. Possible use cases:

• Mimicking code reviews in a specific senior's style.
• Generating ticket responses in corporate tone.
• Simulating chat discussions for testing.

Versioning changes allows experimentation without losing the base model. The author emphasizes ethical aspects but focuses on technical implementation.

Key Points:

• Dual structure (Work Skill + Persona) ensures precise imitation of technical and behavioral patterns.
• Support for retraining with rollback minimizes model degradation risks.
• Adaptation to Chinese corporate cultures with specific tags.
• AgentSkills standard for compatibility with Claude Code and OpenClaw.
• Demo status: awaiting feedback on stability and edge cases.

Prospects for Use in Teams

The tool is relevant for middle/senior developers in distributed teams. Features include automating routine tasks while preserving team style, onboarding via mentor clones, and stress-testing communication. Requires careful handling of data privacy.

— Editorial Team