COLLEAGUE.SKILL: Trace-to-Skill Distillation

• COLLEAGUE.SKILL is an automated system that distills person-grounded skills from diverse work traces into structured, versioned dual-track artifacts.
• It employs a dual-track architecture where the capability track captures expertise and decision heuristics while the bounded behavior track defines communication style and corrections.
• The system incorporates lifecycle management with features like versioning, rollback, and correction handling to adapt expertise while mitigating privacy and bias risks.

Searching arXiv for the cited papers and closely related work on COLLEAGUE.SKILL, skill distillation, and skill retrieval. COLLEAGUE.SKILL is an automated trace-to-skill distillation system for generating person-grounded AI skills via expert knowledge distillation. It addresses a setting in which actionable knowledge associated with a person or role is embedded in heterogeneous traces rather than written as explicit instructions. Given materials from a target person or role, the system 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 is designed to be inspectable, correctable, portable across skill-compatible agent hosts, and explicitly bounded: it represents selected evidence-grounded expertise and style rather than claiming to simulate or replace the target person (Zhou et al., 29 May 2026).

1. Definition and conceptual scope

The central formalization in COLLEAGUE.SKILL treats person-grounded skill generation as an artifact construction problem. Given a lightweight profile $p$, a source scope $c$, and a set of source materials $D = \{d_1, \ldots, d_n\}$, the system produces a skill package

$S = (A, M, L),$

where $A$ is a set of generated files, $M$ is machine-readable metadata and installation information, and $L$ is lifecycle state including versioning and rollback history (Zhou et al., 29 May 2026).

The package is required to satisfy five operational properties: portability, inspectability, composability, correctability, and governability. Portability means that a skills-compatible agent can load it through standard mechanisms. Inspectability means that users can read rules, examples, limitations, and metadata directly. Composability means that the full skill, the work-only track, and the persona-only track can be invoked separately. Correctability means that natural-language feedback or new evidence can update the artifact while preserving prior state. Governability means that metadata supports deletion, sharing decisions, and safety review (Zhou et al., 29 May 2026).

The notion of “person-grounded” is deliberately bounded. The system grounds instructions and patterns in evidence from traces of a specific person or role, but it does not present the result as identity replacement. This boundary is foundational to the design: the artifact is a technical package with explicit evidence scope, usage limits, and lifecycle controls rather than an unrestricted impersonation layer (Zhou et al., 29 May 2026).

2. Dual-track package architecture

The defining representational feature of COLLEAGUE.SKILL is its dual-track structure. The capability track encodes what the person does and how the person reasons: responsibilities, workflows, procedures, technical standards, review criteria, decision heuristics, escalation rules, mental models, and lessons learned. The bounded behavior track encodes how the agent should communicate and behave under explicit constraints: tone, level of detail, interaction posture, boundaries, and correction history. In the implementation, these are materialized as work.md and persona.md, and then rendered into runtime entrypoints (Zhou et al., 29 May 2026).

The combined runtime artifact is SKILL.md, which embeds both tracks and includes operating rules for their joint application. Part A carries the capability track; Part B carries the behavior track. The file also includes standard skill frontmatter such as name, description, and user-invocable: true. This separation is intended to prevent capability and behavior from collapsing into a single opaque persona prompt (Zhou et al., 29 May 2026).

File	Role	Main contents
SKILL.md	Combined runtime entrypoint	Frontmatter, Part A capability, Part B behavior
work.md	Editable capability source	Procedures, standards, heuristics, task patterns
persona.md	Editable behavior source	Style, interaction posture, boundaries, correction log
work_skill.md	Capability-only runtime entrypoint	Generated from work.md
persona_skill.md	Behavior-only runtime entrypoint	Generated from persona.md
manifest.json	Installation and gallery metadata	Entrypoints, compatible runtimes, slash commands, toolchain metadata
meta.json	Lifecycle metadata	Schema version, provenance, lifecycle version, correction count, rollback history pointers

This package design has close affinities with other portable skill formats. In telecommunications operations, SKILLS represents procedural guidance as a SKILL.md document that encodes workflow logic, API patterns, business rules, constraints, and required outputs, and evaluates such documents as external, model-agnostic guidance for tool-using agents (Brett, 16 Mar 2026). A plausible implication is that COLLEAGUE.SKILL extends the same packaging intuition from domain workflows to person-grounded expertise and interaction style.

3. Trace-to-skill distillation workflow

The distillation pipeline begins with heterogeneous traces. Supported evidence sources include work communication such as Feishu, DingTalk, Slack, WeChat chat logs and email archives; work artifacts such as design documents, incident reports, code review comments, Markdown documents, PDFs, and screenshots; public materials such as interviews, essays, speeches, transcripts, subtitles, and research notes; and, for the relationship preset, private interaction histories such as chats and emails under strict consent assumptions (Zhou et al., 29 May 2026).

At a high level, the workflow maps normalized traces into a package: $$f_{\text{distill}} : (p, c, D) \mapsto S = (A, M, L).$$ The pipeline has six stages. First, the user selects a preset and provides an alias, short profile, source scope, and source materials. Second, collectors and parsers normalize those materials into a local knowledge directory. Third, analyzers extract recurring workflows, task patterns, decision heuristics, standards, mental models, communication styles, and interaction rules. Fourth, builders organize these findings into structured Markdown outlines for work.md and persona.md. Fifth, a shared writer assembles the full artifact set, normalizes metadata, sets schema version v3, and generates runtime entrypoints. Sixth, the resulting package is either installed locally or prepared for controlled sharing (Zhou et al., 29 May 2026).

The prompts deliberately separate capability extraction from behavior extraction. The capability side emphasizes durable work methods rather than one-off actions. The behavior side emphasizes stable expression patterns, interaction rules, refusals, and boundaries. This separation is important because the paper frames the artifact not as a flat memory dump but as a structured distillation of two different kinds of evidence: procedural expertise and bounded interactional conduct (Zhou et al., 29 May 2026).

4. Correction, versioning, and rollback

COLLEAGUE.SKILL treats initial generation as provisional and centers a correction lifecycle around natural-language feedback. Every package carries lifecycle state including version, update_time, correction_count, and rollback_history. When users inspect SKILL.md, work.md, or persona.md, they can submit corrections such as “She always prioritizes security over UX in API reviews” or “He would push back more strongly when requirements are vague,” after which the system routes the update to the appropriate track and increments the package version (Zhou et al., 29 May 2026).

Capability corrections are applied as targeted patches to work.md. If a patch refers to an existing Level-2 heading, that section is replaced; if it refers to a new heading, it is appended. Behavior corrections are normalized as structured correction records of the form

$\{\text{scene}, \text{wrong}, \text{correct}\},$

which are appended to a correction log in persona.md. After either kind of update, the writer archives the previous state, regenerates SKILL.md, work_skill.md, and persona_skill.md, and records the new lifecycle state (Zhou et al., 29 May 2026).

Rollback is version-level. If versions are $S^0, S^1, \ldots, S^t$, rollback selects an earlier snapshot $c$0 and restores it as current state. This makes the artifact editable in the style of configuration management rather than irreversible prompt mutation (Zhou et al., 29 May 2026).

This lifecycle differs in source and objective from several agent-skill maintenance systems. AutoSkill derives reusable skills from dialogue traces and maintains a per-user skill bank through add, merge, and discard decisions without retraining the underlying model (Yang et al., 1 Mar 2026). CODESKILL learns a management policy that extracts, evolves, merges, and prunes coding skills from agent trajectories to improve a frozen downstream coding policy (Li et al., 25 May 2026). SkillAdaptor performs training-free step-level revise-or-generate updates under explicit qualification checks to prevent harmful skill drift (Yu et al., 31 May 2026). COLLEAGUE.SKILL shares the externalized-skill and versioned-artifact ethos of these systems, but its update target is a person-grounded package distilled from human traces rather than an environment-grounded skill bank distilled from agent execution (Zhou et al., 29 May 2026).

5. Deployment surfaces, presets, and distribution

The generated package is designed for installation in skills-compatible agent hosts including Claude Code, OpenClaw, Codex, and Hermes. Invocation can target the full SKILL.md, the capability-only work_skill.md, or the behavior-only persona_skill.md. This supports different usage modes: a task may require both the target person’s heuristics and style, only the heuristics, or only the bounded interaction pattern (Zhou et al., 29 May 2026).

The system organizes its workflow through presets. The colleague preset focuses on internal work traces and emphasizes practical work practice, review judgment, and role responsibilities. The celebrity or public-figure preset restricts evidence to public materials such as first-person writings, interviews, documented decisions, transcripts, and subtitles, and adds quality checks for grounding and evidence limits. The relationship preset is local-first and emphasizes interaction patterns, consent assumptions, retention limits, and non-public defaults. Technically, presets are configurations over the same core workflow, differing in source boundary, storage root, command aliases, prompt bundles, and optional research or safety tools (Zhou et al., 29 May 2026).

COLLEAGUE.SKILL is also presented as an open-source ecosystem rather than only a local pipeline. The repository https://github.com/titanwings/colleague-skill had approximately 18.5k GitHub stars as of 2026-05-28. The public gallery listed 215 skills, 55 meta-skills, and 165 contributors, with more than 100k cumulative stars across skill cards (Zhou et al., 29 May 2026).

As skill libraries grow, routing becomes a systems problem rather than only a packaging problem. R3-Skill formulates this explicitly as query-conditional skill routing and shows that skill retrieval differs from document retrieval because top-$c$1 correctness depends on both per-skill relevance and query-conditional compatibility among the retrieved skills. Its two-stage pipeline, R3-Embedding plus R3-Reranker, attains Hit@1 = 0.7714, NDCG@10 = 0.8327, and Set-Compat = 0.3525 on its bilingual benchmark (Wang et al., 2 Jun 2026). This suggests that large COLLEAGUE.SKILL galleries may eventually require compatibility-aware routing layers in addition to portable package formats.

6. Position in the skill-systems literature and open issues

COLLEAGUE.SKILL sits at the intersection of persona systems, memory systems, and skill frameworks, but its distinctive contribution is to define an end-to-end trace-to-skill workflow for person-grounded artifacts. Persona prompts are typically hand-written and difficult to version or correct. Retrieval-augmented memory systems store fragments in hidden indexes rather than surfacing an explicit artifact. Skill frameworks define how to package a skill but not how to derive one from real human traces. COLLEAGUE.SKILL combines evidence collection, dual-track distillation, artifact generation, correction handling, version management, and multi-host deployment into a single workflow (Zhou et al., 29 May 2026).

Related work clarifies the broader landscape. AutoSkill externalizes reusable behaviors from user interactions into versioned skill tuples and SKILL.md artifacts, emphasizing lifelong personalization without retraining (Yang et al., 1 Mar 2026). OpenClaw-Skill constructs a tree of reusable skills from collective agent trajectories and trains agents to choose among them through skill-augmented SFT and collective skill reinforcement learning (Lin et al., 15 Jun 2026). SKILLS shows that portable SKILL.md documents can materially improve telecom-operations agents working through real TM Forum APIs, with consistent lift across model conditions and especially large gains on complex scenarios (Brett, 16 Mar 2026). Taken together, these systems suggest a continuum from task-grounded skills distilled from agent trajectories to person-grounded skills distilled from human traces; COLLEAGUE.SKILL occupies the latter end of that continuum (Zhou et al., 29 May 2026).

The paper is explicit about limitations. It reports no formal task-performance benchmarks, no A/B evaluations, and no behavioral-fidelity metrics; its claims are primarily at the artifact and systems-design level. It also identifies several risks: misrepresentation of expertise or style, bias and privacy risks in source traces, overfitting to limited evidence, and staleness as the real person’s practice evolves. Recommended mitigations include explicit participation and consent, scoped collection and retention limits, straightforward deletion paths, human review before deployment or sharing, and clear disclaimers—especially for public-figure and relationship presets (Zhou et al., 29 May 2026).

These constraints are substantive rather than incidental. Because the system packages work heuristics, mental models, style, and correction history into a portable artifact, governance is part of the technical design. A plausible implication is that COLLEAGUE.SKILL should be understood less as a prompt-engineering convenience than as infrastructure for representing bounded human expertise in agent-readable form: versioned, inspectable, reversible, and institutionally governable (Zhou et al., 29 May 2026).