215 Skills, 165 Contributors, No Fidelity Test

Notes on COLLEAGUE.SKILL: Automated AI Skill Generation via Expert Knowledge Distillation (arXiv:2605.31264) — Tianyi Zhou, Dongrui Liu, Lei Yuan, Jing Shao, Xia Hu · May 2026

Note published July 25, 2026 · written by SkillFed’s research pipeline from the paper above · how these notes are made

AI-assisted notes · reviewed by SkillFed Skill evolution

COLLEAGUE.SKILL turns raw traces of a person — chat logs, work documents, email, screenshots, public research material, subtitles, plus a lightweight description — into a skill package through person-grounded trace-to-skill distillation: not a simulation of the person, but a bounded, editable artifact extracted from them. Each package splits into two coordinated tracks: a capability track for practices, mental models, and decision heuristics, and a behavior track for communication style, interaction rules, and correction history. The output is five content files — a combined SKILL.md plus independently invokable work.md/persona.md and work_skill.md/persona_skill.md — wrapped in manifest.json and meta.json for install and lifecycle state, on schema version 3. It installs across four agent hosts (Claude Code, OpenClaw, Codex, Hermes) and ships three domain presets: colleague for private workplace material, celebrity/public-figure for public evidence under source boundaries, and relationship for personal traces under consent and local-control rules.

Corrections run through natural language and land as either a Markdown patch or a normalized scene/error/correction record — the system's correction lifecycle — with every revision versioned and rollback-able. By May 28, 2026 the open-source deployment had pulled in roughly 18,500 GitHub stars; the public gallery listed 215 skills and 55 meta-skills from 165 contributors, with cumulative star counts across skill cards topping 100,000. Those are adoption numbers, not accuracy numbers. The paper validates the artifact — inspectable, correctable, installable — not whether a distilled skill actually preserves the source person's judgment; that question gets deferred to future evaluation work.

Key numbers

GitHub stars (repository)	~18,500
Skills listed in public gallery	215
Gallery contributors	165
Meta-skills listed	55
Cumulative gallery-card stars	>100,000

What this means for skill authors

Strip away the person-grounded framing and what's left worth borrowing is the split itself: capability vs. behavior as separately invokable artifacts. Most skill authors fuse "what to do" and "how to sound doing it" into one prompt blob; here work_skill.md and persona_skill.md ship as standalone entrypoints, so a downstream agent can pull the decision heuristics without inheriting a specific communication style, or the reverse. That's a composability pattern any skill format benefits from, not just ones built from someone's chat history.

The correction lifecycle is the more durable contribution. A skill drifts wrong, someone corrects it in plain language, the system emits a versioned patch and keeps a record of what changed and why, and the whole thing can roll back. That's close to what any catalog needs once skills are out in the field and occasionally wrong — version numbers and correction counts as visible lifecycle metadata give a reviewer something to check beyond just reading current content.

The gap sits under the headline numbers. 215 skills, 165 contributors, 100k+ stars describe adoption of a packaging format, not evidence that any given skill reproduces the judgment it claims to distill — the authors are explicit that behavioral fidelity is untested and left to future benchmarks. That's exactly the distinction dual-tool scanning and a vetted catalog exist to draw: a skill that installs cleanly and passes structural checks has cleared the bar this paper sets, not the higher one users will assume it cleared.

Skills related to this research

drawio-skill drawio-skill creates polished diagrams—architecture, UML, ER, flowcharts, BPMN, and more—from text descriptions or code imports. It generates .drawio XML and exports locally via the draw.io desktop CLI to PNG, SVG, PDF, or JPG, with optional embedding for editability. Use it for precise, styled diagrams with rich shape libraries and swimlanes; for casual sketches or Markdown-embedded diagrams, consider mermaid or excalidraw instead.★ 6,723 Portfoliosyncing PortfolioSyncing imports and synchronizes broker portfolio exports into Google Sheets, keeping your position data current across devices. It automates Fidelity CSV parsing, validates quantity and cost basis changes, and protects critical formulas while updating holdings.★ 315 axiom-profile-performance axiom-profile-performance equips AI coding assistants with specialized capabilities for measuring and optimizing app performance on Apple platforms. It enables headless profiling workflows that automatically gather execution traces and surface performance bottlenecks without manual instrumentation. Built for modern Swift development, this skill integrates seamlessly with Claude Code and other AI-native development environments.★ 1,095 faceless-explainer faceless-explainer transforms written content into animated explainer videos by generating visuals frame by frame. It handles topic explanations, concept breakdowns, and how-tos without requiring website capture or existing footage. Work through a structured workflow—setup, brief extraction, design system selection, storyboarding, frame composition, and final rendering—orchestrating each step to build your video.★ 576

Related notes

References

Zhou, Liu, Yuan, Shao, Hu. "COLLEAGUE.SKILL: Automated AI Skill Generation via Expert Knowledge Distillation." arXiv:2605.31264 (2026).
Yao, Zhao, Yu, Du, Shafran, Narasimhan, Cao. "ReAct: Synergizing Reasoning and Acting in Language Models." arXiv:2210.03629 (2023).
Schick, Dwivedi-Yu, Dessì, Raileanu, Lomeli, Zettlemoyer, Cancedda, Scialom. "Toolformer: Language Models Can Teach Themselves to Use Tools." arXiv:2302.04761 (2023).
Shinn, Cassano, Berman, Gopinath, Narasimhan, Yao. "Reflexion: Language Agents with Verbal Reinforcement Learning." arXiv:2303.11366 (2023).
Wu, Bansal, Zhang, Wu, Zhang, Zhu, Li, Jiang, Zhang, Wang. "AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation." arXiv:2308.08155 (2023).