Source-linked AI summary
COLLEAGUE.SKILL: Automated AI Skill Generation via Expert Knowledge Distillation
Tianyi Zhou, Dongrui Liu, Leitao Yuan, Jing Shao, Xia Hu
TL;DR
Turning heterogeneous human traces into usable, bounded expertise remains difficult because existing systems lack an end-to-end, inspectable workflow. COLLEAGUE.SKILL distills selected traces into portable, correctable skill artifacts, illustrating a practical alternative to opaque prompts or faithful-person simulations.
Problem
Existing memory, persona, and skill systems do not provide an end-to-end workflow for distilling heterogeneous human traces into inspectable, correctable, agent-usable skills.
Method
COLLEAGUE.SKILL normalizes traces, extracts capabilities and bounded interaction styles, and renders versioned packages supporting inspection, correction, rollback, installation, and optional distribution.
Results
The system demonstrates that selected human traces can be distilled into portable, inspectable skills encoding capabilities, mental models, behavior constraints, and correction history.
Takeaways & Limitations
Person-grounded distillation is best represented as editable artifacts that users can read, revise, install, withhold, and delete rather than opaque prompts.
Takeaways & Limitations
The paper leaves source matching, task performance, emotional safety, and user-trust calibration open, with outcomes dependent on source quality, extraction, model behavior, and human review.
Abstract
from arXiv · showhide
LLM agents are increasingly expected not only to complete isolated tasks, but also to carry bounded representations of human expertise, judgment, and interaction style. Building such person-grounded agents remains difficult because actionable knowledge associated with a person or role is usually embedded in heterogeneous traces rather than written as clean instructions. Existing memory and persona systems capture fragments of this evidence, while skill frameworks provide portable packaging formats; however, there is no end-to-end workflow for distilling these traces into inspectable, correctable, and agent-usable skills. We present an automated trace-to-skill distillation system for generating person-grounded AI skills via expert knowledge distillation. Given materials from a target person or role, COLLEAGUE.SKILL produces a versioned skill package with two coordinated tracks: a capability track for practices, mental models, and decision heuristics, and a bounded behavior track for communication style, interaction rules, and correction history. The package can be inspected, invoked, updated through natural-language feedback, rolled back, installed across agent hosts, and optionally prepared for controlled distribution. We describe the artifact contract, generation workflow, correction lifecycle, deployment surface, and domain presets implemented in the open-source system. At the time of writing, the public repository has approximately 18.5k GitHub stars; the gallery lists 215 skills from 165 contributors and more than 100k cumulative stars across listed skill cards. The system illustrates how person-grounded skills can be represented as portable, correctable packages rather than opaque prompts or hidden memories.
1 Introduction
COLLEAGUE.SKILL addresses the under-specified problem of creating bounded, person-grounded agent skills from heterogeneous traces rather than existing instruction manuals. It frames the result as a portable, inspectable, correctable artifact with coordinated capability and behavior components, installation, correction, rollback, and governance support.
- Motivation: LLM agents are increasingly expected to preserve bounded parts of a person’s expertise, memory, judgment, or interaction style for reusable work and interaction.Examples include review judgment, decision heuristics, mental models, and interpersonal patterns.
- Problem: Person-grounded knowledge is dispersed across heterogeneous traces, including code comments, incident notes, chat decisions, interviews, speeches, essays, and public decisions.These traces are not typically organized as clean instruction manuals.
- Approach: COLLEAGUE.SKILL distills selected traces into a bounded package of capabilities, mental models, communication constraints, examples, and usage boundaries.The system treats traces as evidence for a portable skill rather than a hidden memory store or unrestricted simulation of an individual.
- Artifact: The pipeline produces a capability track, a bounded behavior track, metadata, host installers, and version state from heterogeneous human traces.The capability track covers reusable practices and reasoning resources, while the behavior track constrains communication and interaction patterns.
- Lifecycle: The artifact workflow supports collection, skill rendering, multi-host installation, natural-language correction, rollback, portability, inspectability, composability, and governance.These requirements position person-grounded trace-to-skill distillation as an explicit artifact-construction problem.
2 Problem Formulation
COLLEAGUE.SKILL formulates person-grounded skill generation as producing a bounded, skills-compatible artifact from a profile, source scope, and source materials. The target is an inspectable, portable, composable, correctable, and governable package rather than a hidden behavioral model of a person.
- Scope: Person-grounded skills ground instructions in evidence about a person or role while imposing explicit source, usage, and governance constraints.The colleague setting is the primary instance, but the artifact form also covers public mental models and private interaction patterns under different constraints.
- Artifact definition: Given a lightweight profile, source scope, and source materials, the system produces a skill package with generated files, machine-readable metadata, installation information, and lifecycle state.Lifecycle state includes version, update time, correction count, and rollback history.
- Operational properties: The package operationalizes five properties: portable, inspectable, composable, correctable, and governable.These properties support ordinary skill loading, pre-use review, separate entrypoints, feedback-driven updates that preserve prior state, and safety-oriented source and metadata controls.
- Domain instantiation: Domain-specific source materials and controls vary: work traces support colleague skills, public first-person evidence supports public-figure skills, and private relationship traces require consent and local control.The source examples include design documents, code-review comments, chat decisions, incident notes, and long-form interviews.
- Claim boundary: The formulation claims trace transformation into a skills-compatible artifact with explicit files, metadata, entrypoints, correction records, and lifecycle operations, not faithful behavioral cloning.This scope enables inspection of structure, source boundaries, update behavior, and deployment compatibility.
3 COLLEAGUE.SKILL System Overview
COLLEAGUE.SKILL provides a shared, trace-to-artifact pipeline that extracts capability and bounded interaction evidence into portable person-grounded skills. Domain presets and independently invokable artifact tracks adapt this workflow to different source, consent, and usage requirements.
- Architecture: The pipeline normalizes work documents, reviews, interviews, writings, or interaction records, then analyzes them for durable capability, mental models, and bounded interaction style.Collectors and parsers create local knowledge directories before analyzers and builders produce structured artifacts.
- Domain presets: Three presets—colleague, celebrity, and relationship—reuse one artifact workflow while varying source boundaries, storage roots, command aliases, prompts, and optional tooling.Future presets can be added through configuration and prompt design rather than a new program.
- Dual representation: The dual representation separates reusable expert judgment from bounded behavior constraints, expression preferences, interaction rules, and correction records.The capability track covers responsibilities, workflows, standards, review criteria, heuristics, and lessons from past work; persona.md stores the narrower behavior track.
- Runtime surface: Generated artifacts expose full, capability-only, and persona-only entrypoints, making factual knowledge, procedural judgment, and surface behavior separately inspectable and invocable.The combined SKILL.md embeds capability as Part A and behavior as Part B, while split entrypoints expose each track independently.
- Artifact contract: The shared writer emits a versioned schema and Agent Skills-compatible SKILL.md entrypoint with progressive disclosure of detailed instructions at invocation.The current implementation uses schema version 3 and also renders editable source documents plus installation, gallery, and lifecycle metadata.
4 Generation and Evolution Workflows
COLLEAGUE.SKILL turns heterogeneous person- or role-grounded source material into inspectable skill artifacts through separate capability and behavior tracks. Generated skills evolve through natural-language corrections, versioned regeneration, and domain-specific evidence and privacy constraints.
- Creation workflow: Creation ingests person- or role-grounded materials through supported collectors and imports, then specializes prompts with application presets.Sources include Feishu, DingTalk, Slack, WeChat SQLite exports, email archives, PDFs, screenshots, Markdown, and direct paste.
- Two-track generation: The capability track extracts methods, heuristics, and mental models, while the behavior track extracts bounded expression and interaction patterns into inspectable Markdown artifacts.The writer packages the structured Markdown into the defined artifact contract, keeping capability and behavior claims separate at file level.
- Correction lifecycle: Natural-language corrections update expert-work sections through Markdown patches or behavior through normalized correction records.Matching level-2 headings replace sections; unmatched patch sections are appended.
- Correction lifecycle: Versioned evolution archives the current artifact, applies changes, increments the lifecycle version, regenerates derived artifacts, and supports listing, backup, rollback, and archive cleanup.Corrections and patches create new versions while preserving rollback points.
- Domain presets: Domain presets extend the workflow with public-source evidence controls for celebrities and stronger consent, retention, access-control, and sharing assumptions for relationships.The celebrity extension prioritizes first-person works, long-form interviews, documented decisions, and marked inferences; it can downgrade confidence when evidence is thin but does not certify factual truth.
5 Deployment and Community Ecosystem
COLLEAGUE.SKILL is deployed as an open-source artifact pipeline with a public site and gallery supporting local use, installation across agent hosts, and controlled community sharing. Its deployment surface also exposes public ecosystem counters while distinguishing distribution signals from task-performance or behavioral-fidelity metrics.
- Deployment surface: The open-source repository, public site, and gallery document workflows, installation options, supported sources, example outputs, and downstream skill sharing.Generated skills may remain local, be installed into an agent host, or be submitted as shareable packages when users have publication rights.
- Community ecosystem: 215 skills, 55 meta-skills, and 165 contributors were listed by public counters observed on 2026-05-28.These counts summarize repository activity and gallery scale rather than task performance, behavioral fidelity, or adoption quality.
- Artifact lifecycle: Deployment reframes COLLEAGUE.SKILL from single-prompt construction into a portable artifact pipeline whose skills can be used, corrected, installed, or withheld from public sharing.The public site and gallery are presented as part of the artifact story, showing movement from local use to controlled installation and appropriate community sharing.
6 Application Cases
The application cases show the shared trace-to-skill workflow across workplace, public-figure, and interpersonal domains. They are design-oriented examples rather than claims of behavioral equivalence, with domain-specific evidence boundaries and safeguards.
- Overview: The cases apply a shared trace-to-skill workflow across different domains as design-oriented examples, not claims of behavioral equivalence.The examples cover workplace, public-figure, and sensitive interpersonal settings.
- Colleague skill: Workplace skills distill private or enterprise materials into reusable work practice centered on review criteria such as authentication and input validation.The supplied passage also names rate limiting, response schema, and sensitive-data exposure as review criteria.
- Celebrity skill: Public-figure skills use a six-dimensional research pass and quality checks to emphasize mental models, citations, and explicit boundaries.Generated skills should identify thin evidence, avoid presenting themselves as the actual person, and remain distinct from enterprise-trace workplace skills.
- Relationship skill: Relationship skills represent interaction patterns as local, editable state while addressing emotional overattachment, non-consensual simulation, and private-chat misuse.Deployment should prioritize local ownership, deletion, clear disclaimers, and opt-in sharing; the case is presented as an extension capability rather than an endorsement of unconstrained use.
7 Related Work
Related work externalizes agent capabilities through reusable skills, execution feedback, package specifications, memory, personalization, and role-playing frameworks. COLLEAGUE.SKILL is positioned within these directions while adopting an emerging package format.
- Agent skills and reusable capabilities: Recent agent systems externalize capabilities through reasoning-action interleaving, learned API use, feedback-driven revision, and configurable multi-agent conversations.Examples include ReAct, Toolformer, Reflexion, Self-Refine, and AutoGen.
- Agent skills and reusable capabilities: Agent Skills specifies SKILL.md-centered directories with optional scripts, references, and assets loaded through progressive disclosure.The format is described as an emerging infrastructure layer for agents, alongside concerns about redundancy, marketplace skew, and safety risks.
- Skill libraries and skill synthesis: Skill libraries support reusable behavior, while newer systems distill or refine skill knowledge from execution trajectories.Voyager stores executable code skills, and SkillX constructs hierarchical skills refined through execution feedback.
- Memory, personalization, and role-playing agents: Memory and personalization provide continuity through retrieval stores, context managers, model behavior, user histories, preferences, and personalized action spaces.Retrieval-augmented generation connects parametric generation with non-parametric memory, while LaMP and PersonaAgent study personalization.
8 Discussion
The discussion frames COLLEAGUE.SKILL as a bounded, inspectable artifact distilled from grounded traces rather than an identity replacement. Its research contribution lies in the workflow and product surface that support accountability, correction, portability, and controlled reuse while leaving behavioral fidelity as an open question.
- Grounded traces, not identity replacement: COLLEAGUE.SKILL distills bounded representations of knowledge, judgment, and interaction style without claiming to reproduce the person.The target includes evidence weighting, risk detection, trade-off explanation, refusal of bad requests, contextual communication, and documented interaction rules.
- Why the workflow matters: The system treats trace-to-skill distillation as an auditable file workflow spanning creation, inspection, invocation, correction, rollback, deletion, installation, and optional distribution.These operations enable generated skills to be audited, repaired, withheld, or shared under explicit controls.
- Reusable person-grounded artifacts: The colleague, celebrity, relationship, and gallery settings instantiate one thesis: reusable person-grounded artifacts should preserve visible evidence boundaries and limits rather than simulate unbounded identities.The proposed ecosystem includes security-review, product-decision, research-mentor, public-thinker mental-model, and private-interaction skills.
- Behavioral fidelity frontier: The paper makes artifact-level claims about package format, generation and update workflows, correction and rollback state, multi-host support, and coverage across four settings, not faithful reproduction or downstream improvement.Behavioral fidelity and effects on downstream work require human evaluation and remain unclaimed here.
- Productization as a research constraint: Productization is part of the contribution because installers, manifests, gallery metadata, rollback state, and deletion paths make skills legible across tools, teams, and sharing contexts.These surfaces also provide handles for future comparisons of source scopes, correction records, and invocation behavior.
9 Limitations and Responsible Deployment
COLLEAGUE.SKILL presents person-grounded skills as editable artifacts rather than faithful simulations, identity substitutes, or consent proxies, while leaving several deployment questions open. Responsible use requires participation, scoped collection, access controls, retention limits, non-mandatory use, and separate review of consent, lawful source use, and redaction.
- Scope and open questions: The system does not claim faithful simulation, identity substitution, or consent proxying for the grounded person.It documents the artifact format, workflow, implementation, and public deployment surface instead.
- Scope and open questions: Source quality, extraction quality, model behavior, and human review will shape real-world deployments.Source matching, task performance, emotional safety, and user-trust calibration remain open.
- Responsible deployment: Responsible deployment requires explicit participation, scoped source collection, access controls, retention limits, and non-mandatory use.Local-first, inspectable, and versioned design provides governance affordances, but lawful source use, consent, and full redaction require separate review.
- Responsible deployment: Gallery publication should remain opt-in and include submitter attestation, review, takedown, source-boundary labels, and visible disclaimers.These safeguards apply especially to celebrity, public-figure, or relationship-oriented cases.
10 Conclusion
COLLEAGUE.SKILL frames person-grounded agents as portable, inspectable skills distilled from selected human traces rather than attempts to recreate people. It also proposes bounded digital-double packages and benchmarks centered on useful judgment, provenance, consent, and visible failure modes.
- Core contribution: The system distills selected human traces into portable, inspectable skills encoding capabilities, mental models, behavior constraints, and correction history.The colleague setting is the concrete starting point.
- Scope: The artifact model extends beyond colleagues to celebrity, relationship, and gallery components.These components demonstrate broader applicability of the same model.
- Future evaluation: Future digital-double benchmarks should test whether bounded packages preserve useful judgment while exposing provenance, consent, and failure modes.The proposed framing emphasizes explicit evidence, rights, correction semantics, and distribution choices rather than open-ended impersonation.
Agent Skills. Agent skills specification. https://agentskills.io/specification, 2026b. Accessed
This section lists related work spanning agent skills, retrieval, tool use, personalization, role-playing, reflection, and multi-agent systems. It also cites recent approaches to skill synthesis, skill evolution, and personalized agents.
- Agent Skills: Related work includes agent-skill specifications, Claude skills, retrieval-augmented generation, and data-driven analyses of agent skills.These references cover both skill interfaces and knowledge retrieval for language-model agents.
- Agent Skills: The bibliography cites Toolformer, ReAct, AutoGen, AgentBench, Voyager, and SkillGen for tool use, agent evaluation, multi-agent interaction, and skill synthesis.Together, these works represent infrastructure, benchmarking, embodied exploration, and verified inference-time skill construction.
- Agent Skills: The references also include Self-Refine, Reflexion, SkillX, and AutoSkill, covering iterative feedback, verbal reinforcement, skill knowledge bases, and lifelong skill self-evolution.These approaches emphasize updating or constructing reusable agent capabilities through feedback and experience.
- Agent Skills: Personalization and role-grounded behavior are represented by LaMP, Character-LLM, RoleLLM, PersonaAgent, and SOTOPIA.The cited works address personalization, role-playing, persona conditioning, and interactive social intelligence.