Overview
Paper: COLLEAGUE.SKILL: Automated AI Skill Generation via Expert Knowledge Distillation Authors: Tianyi Zhou, Dongrui Liu, Leitao Yuan, Jing Shao, Xia Hu (Shanghai AI Laboratory) Paper: https://arxiv.org/abs/2605.31264 Code: https://github.com/titanwings/colleague-skill (18.5k stars) Project page: https://titanwings.github.io/colleague-skill-site/ Community: 215 skills / 165 contributors / 100k+ cumulative gallery stars
Key points
- Problem: LLM agents increasingly need to carry reusable work context—a teammate's code-review standards, an expert's judgment heuristics—but this knowledge is scattered across chat logs, docs, emails and meeting notes. Existing approaches fall short: memory systems capture only fragments; persona systems conflate knowledge, judgment and tone into superficial imitation; skill frameworks provide packaging but no end-to-end pipeline from raw traces to skills.
- Core insight: convert selected human traces into skill packages that are *inspectable, correctable, installable and governable*—rather than hidden memory stores or claims of reproducing a person.
- Dual representation: each skill has a Capability Track (
work.md: work methods, mental models, decision heuristics, technical standards) and a Behavior Track (persona.md: communication style, interaction rules, boundaries, corrections). They can be invoked independently (/{character}-{slug}-workfor pure capability,/{character}-{slug}-personafor pure style) or combined. - Three application presets sharing one workflow: *colleague* (Feishu/DingTalk/Slack/WeChat/email/docs; enterprise governance), *celebrity* (interviews, talks, books; source boundaries, fact-checking, copyright safety), *relationship* (private chats, diaries; informed consent, local control, deletability).
- Artifact structure: a standard skill package containing
SKILL.md,work.md,persona.md,work_skill.md,persona_skill.md,manifest.jsonandmeta.json. Packages are portable (Claude Code, OpenClaw, Codex, Hermes), inspectable, composable, correctable and governable. - Creation workflow: collectors/parsers normalize source materials into a local knowledge catalog → analyzer extracts durable capabilities and interaction patterns → builder renders structured Markdown → writer packages the artifact.
- Correction workflow: natural-language feedback (e.g., "she would push back here") is classified as capability vs. persona, generates a Markdown patch or a normalized correction record
{scene, wrong, correct}, archives the current version, applies the patch and bumps the version. Commands:list,rollback,backup,clean. - Celebrity research toolchain: subtitle download → transcript conversion → research-note merging → quality check, plus a six-dimension research framework (Works, Interviews, Decisions, Expression DNA, External Evaluations, Timeline). The quality checker scans mental-model coverage, limitation notes, expression patterns, internal contradictions, source URLs and copyright-safety signals.
- Deployment: installable to
~/.claude/skills/dot-skill(Claude Code),~/.openclaw/workspace/skills/dot-skill(OpenClaw),~/.codex/skills/dot-skill(Codex), or via script for Hermes, invoked as/dot-skill. - Paper: https://arxiv.org/abs/2605.31264
- Code: https://github.com/titanwings/colleague-skill
- Project page: https://titanwings.github.io/colleague-skill-site/
Relation to prior work
| Direction | Representative work | Difference | |---|---|---| | Skill libraries | Voyager, SkillX, SkillGen, AutoSkill | Synthesize skills from execution traces; COLLEAGUE.SKILL distills from human traces and separates capability from persona | | Memory & personalization | RAG, LaMP, PersonaAgent | Keep representations in retrieval stores; COLLEAGUE.SKILL emits explicit, installable artifacts | | Role-play | Character-LLM, RoleLLM, SOTOPIA | Pursue behavioral cloning; COLLEAGUE.SKILL pursues bounded artifacts, explicitly not identity replication | | Agent Skills | Agent Skills Spec, Claude Skills | Provide packaging format; COLLEAGUE.SKILL provides a full traces-to-skill workflow |
Limitations and responsible deployment
The paper explicitly states: it does not pursue behavioral fidelity; it does not solve identity substitution; output quality depends on source quality (garbage in, garbage out); user corrections may lend unwarranted confidence to contested traces. Responsible deployment requires informed participation, scoped collection, access control and retention limits, non-coercive use, opt-in gallery publishing with verification/review/takedown, and source-boundary disclaimers for celebrity and relationship extensions.
Why it matters
1. A paradigm shift from "simulating a person" (how alike?) to "extracting reusable knowledge" (how useful?)—turning human traces into inspectable rule files, versioned correctable objects, installable artifacts and governable assets. 2. Skills as vehicles for knowledge transfer: when a colleague leaves or an expert retires, their judgment can now persist in a skill package—with sources, boundaries and limitations visible to everyone. 3. Productization as a research constraint: installers, manifests, gallery metadata, rollback state and deletion paths are treated as contributions, enabling reproducible comparison and auditable governance over explicit metadata rather than hidden prompt state.