Key points
- Problem framing. LLM agents are shifting from one-off task execution to carrying reusable work contexts—code-review standards, expert heuristics, mental models, interaction patterns—but this knowledge lives scattered across chats, design docs, emails, and meeting notes. Memory systems capture fragments without structure; persona systems conflate knowledge, judgment, and tone; skill frameworks lack an end-to-end workflow from raw traces to skills.
- Core insight. Convert selected human traces into inspectable, correctable, installable, governable skill packages—not hidden memory stores or identity-replication claims.
- Dual-track representation. Skills split into a Capability Track (work methods, mental models, decision heuristics, technical standards in
work.md) and a Behavior Track (communication style, interaction rules, boundary constraints, correction logs inpersona.md). The split enables three invocation modes:/{character}-{slug}-work,/{character}-{slug}-persona, and/{character}-{slug}(combined). - Three application presets. The same underlying workflow is parameterized by scenario:
- *colleague*: Slack, Feishu, DingTalk, WeChat, email, docs; focuses on work knowledge and enterprise permissions.
- *celebrity*: interviews, talks, writings, subtitles, decision logs; focuses on source boundaries, fact-checking, copyright.
- *relationship*: private chats, diaries, letters; focuses on informed consent, local control, deletion paths.
- Artifact structure. A standard skill package contains
SKILL.md,work.md,persona.md,work_skill.md,persona_skill.md,manifest.json, andmeta.json(lifecycle state: version, correction count, sources). Key properties: Portable (Claude Code, OpenClaw, Codex, Hermes), Inspectable, Composable, Correctable (natural-language patches, versioned rollback), Governable. - Creation workflow. User supplies an alias and source materials (Feishu API, DingTalk browser mode, Slack API, WeChat SQLite export,
.eml/.mbox, PDF, images, Markdown) → collector/parser → analyzer → builder → writer producing the packaged artifact. - Correction workflow. Natural-language feedback (e.g., "he would not say that") is classified as capability or persona, packaged as a Markdown patch or normalized
{scene, wrong, correct}record, versioned with rollback (list,rollback,backup,clean). - Celebrity research toolchain. Six-dimensional research framework: Works, Interviews, Decisions, Expression DNA, External Evaluations, Timeline. Scripts:
download_subtitles.sh,srt_to_transcript.py,merge_research.py,quality_check.py. The checker audits mind-model coverage, limitation labeling, expression patterns, internal contradictions, source URLs, and copyright safety signals. - Ecosystem (as of 2026-05-28). 215 public skills, 55 meta-skills, 165 contributors, 100k+ cumulative gallery stars.
- Representative cases. Andrej Karpathy-style skill emphasizing scaffolding and strict evaluation over raw demos; ByteDance backend-engineer skill enforcing N+1 fixes and standard response formats; relationship skill with conversational pushback.
- Position vs. related work. Distinct from Voyager/SkillX/SkillGen/AutoSkill (skills from execution traces), RAG/LaMP/PersonaAgent (retrieval-based memory), Character-LLM/RoleLLM/SOTOPIA (behavioral cloning), and Agent Skills Spec/Claude Skills (format only). COLLEAGUE.SKILL is the first to provide an end-to-end trace-to-skill workflow with separated capability/persona tracks.
- Stated limitations. Does not pursue behavioral fidelity; does not solve identity substitution; output quality depends on source quality; user corrections may make contested traces appear more certain.
- Responsible deployment requirements. Informed participation, scoped collection, access control, retention limits, non-coerced use, opt-in gallery with submitter verification, moderation, takedown, and source-boundary labels for celebrity/relationship presets.
- Paper: https://arxiv.org/abs/2605.31264
- Code: https://github.com/titanwings/colleague-skill
- Project page: https://titanwings.github.io/colleague-skill-site/