“"when a colleague leaves, their knowledge walks out the door. Can we keep it?" — Tianyi Zhou, ROADMAP.md”
You know that feeling when your most knowledgeable teammate gives their two weeks' notice and you realize you have no idea how they made half the decisions they did? Their tribal knowledge — why they built the auth system that way, their preferred code review tone, their read on which stakeholders to escalate to — never made it into any doc. colleague-skill addresses this gap: it ingests their existing chat logs, emails, and documents before they leave, and distills them into an AI persona that knows their decision patterns and work style. The alternative is either a rushed knowledge-transfer session that captures maybe 20% of what they knew, or months of re-learning what they already figured out.
You run /create-colleague inside Claude Code and answer three questions: the person's alias, their role summary, and a personality snapshot. Then you choose a data source — Feishu auto-collect pulls messages and docs directly, Slack uses the API, WeChat imports from a SQLite export, or you paste text directly. The tool splits everything into two separate files: work.md covering technical systems, workflows, and output format preferences, and persona.md covering a five-layer personality model governing communication style, decision approach, and MBTI-influenced posture. Both merge into SKILL.md, a single file Claude Code loads as a first-class skill you call with /{slug}. If the distillation gets something wrong — type 'He wouldn't say that' — it writes the fix to the exact layer it belongs in without touching the other.
If you are an engineering lead or IC at a Chinese tech company using Feishu or DingTalk, with a team member about to leave and access to their chat history, this tool is directly usable today. Also relevant if you are exploring the Claude Code skills ecosystem and want a concrete multi-source data ingestion example to study. Not recommended for regulated industries (no consent mechanism built in), English-first orgs relying on Slack free tier (90-day history cap per INSTALL.md), or anyone who needs production-grade stability — no tagged releases exist and only 3 contributors are active.
Worth a test run today if you are in a Feishu-heavy org and have a leaving colleague willing to consent — setup is a single git clone and the skill runs entirely inside Claude Code. Treat it as experimental: no releases and 3 contributors means the API surface will change without warning. The privacy architecture — operator's OAuth for P2P chats, no subject consent mechanism — is a structural limitation that makes enterprise rollout legally complicated in most Western jurisdictions until the project addresses it at the design level.
Deep-dive insight, Easy and Pro modes, plus action playbooks — the full breakdown is one tap away.