COLLEAGUE.SKILL: Distill a Person Into an Agent Skill
Shanghai AI Lab dropped COLLEAGUE.SKILL on HuggingFace Daily Papers, a pipeline that distills heterogeneous human traces (chat logs, docs, email, screenshots, interviews) into a portable agent skill package. They're insistent it's not 'persona impersonation'. Bounded artifacts that capture how a specific colleague does code review, weighs incident severity, or refuses bad PRs.
The construction is unusually clean. Two tracks: capability (heuristics, mental models, review checklists) and behavior (interaction style, correction history). Output is a SKILL.md + work.md + persona.md package with manifest and lifecycle state. You can roll back, patch sections via natural-language feedback ('he would push back here'), install across Claude Code, OpenClaw, Codex, Hermes. The repo titanwings/colleague-skill sits at ~18.5K stars with a gallery of 215 skills from 165 contributors.
The interesting move is generalizing past colleague skills. They use the same pipeline for celebrity skills (mental models of public thinkers from interviews) and relationship skills (private interaction patterns kept local-only with consent). The 'person-grounded artifact, not behavioral cloning' framing is the right one to keep this useful and not gross.
The deeper thesis: skills are eating fine-tuning. If you want an agent that thinks like a senior reviewer, you don't fine-tune a model on their code review history. You distill their judgment into a versioned, inspectable, rollback-able file that any agent host can load. That's a fundamentally different distribution model.
Paper: https://huggingface.co/papers/2605.31264 | Repo: https://github.com/titanwings/colleague-skill
← Back to all articles
The construction is unusually clean. Two tracks: capability (heuristics, mental models, review checklists) and behavior (interaction style, correction history). Output is a SKILL.md + work.md + persona.md package with manifest and lifecycle state. You can roll back, patch sections via natural-language feedback ('he would push back here'), install across Claude Code, OpenClaw, Codex, Hermes. The repo titanwings/colleague-skill sits at ~18.5K stars with a gallery of 215 skills from 165 contributors.
The interesting move is generalizing past colleague skills. They use the same pipeline for celebrity skills (mental models of public thinkers from interviews) and relationship skills (private interaction patterns kept local-only with consent). The 'person-grounded artifact, not behavioral cloning' framing is the right one to keep this useful and not gross.
The deeper thesis: skills are eating fine-tuning. If you want an agent that thinks like a senior reviewer, you don't fine-tune a model on their code review history. You distill their judgment into a versioned, inspectable, rollback-able file that any agent host can load. That's a fundamentally different distribution model.
Paper: https://huggingface.co/papers/2605.31264 | Repo: https://github.com/titanwings/colleague-skill
Comments