Distilly, formerly colleague-skill: turning messages and documents about a person into an installable Agent Skill

Colleague Skill became Distilly, and the old names still circulate

The repository titanwings/colleague-skill now presents itself as Distilly. The README keeps the former name, Colleague Skill, for search continuity and project history, but since an update on 2026-08-24 the creator is named Distilly end to end. The package.json reflects the new identity: the package is @titanwings/distilly, its repository field points to github.com/titanwings/distilly, and it exposes a distilly command from bin/distilly.mjs. The default branch carries a third name, dot-skill.

For readers this naming drift has a practical side. Links to colleague-skill, distilly and the dot-skill branch all point at the same project, and older blog posts or forum threads will describe the Claude Code only colleague version rather than the current multi-host creator. The canonical creator Skill is named distilly, and the project asks you to install it in a directory named distilly. If you copy instructions from an older guide that uses the colleague-skill folder name, the Skill you end up with is not the one current documentation assumes.

One creator Skill, three person families

Distilly describes itself as the person-modeling layer for agents. The pipeline is stated as messages, documents, interviews and public sources going into Distilly, which outputs a Person Profile that an Agent or compatible Bot then uses. You supply source files plus your own description of the person, and the creator builds a profile from observable experience, decision patterns, expression and ways of working. The current release packages each profile as an Agent Skill so supported hosts can install and invoke it.

The creator handles three families with one workflow. The colleague family covers coworkers, mentors, teammates and upstream or downstream partners. The relationship family covers exes, partners, parents, friends and close family. The celebrity family covers public figures, creators, public voices and fictional characters. The README also notes you can model yourself. Each family has its own source collection strategy, its own analysis dimensions and its own Person Profile structure, so picking the right family at the start decides what the profile will even try to capture.

The colleague family builds a Work Skill plus a Persona

The colleague case is where the project started, and it produces two parts. A Work Skill holds technical standards and workflows derived from the person's sources, and a Persona holds their expression and workplace behavior. The pitch is aimed at the moment a colleague quits, a mentor graduates or a teammate transfers and takes their playbook and context with them.

Collection for this family supports Lark, DingTalk and Slack. Those are the workplace platforms named for automated gathering; other chat tools are not listed for collection, so their exports would have to come in as plain documents. The Python dependencies reflect this split: requirements.txt holds only requests>=2.28.0 as the core dependency needed by every Distilly tool, while collector specific extras are moved into requirements-optional.txt. If you only plan to feed in documents you already have, you do not need the optional collector packages; if you want to pull from Lark, DingTalk or Slack, install the optional set and expect to configure access to those workspaces yourself.

Relationship and celebrity profiles draw on different evidence

The other two families are not variations of the colleague template. For relationships, Distilly organizes expression patterns, emotional triggers, conflict patterns and repair patterns found in the sources into a reusable Persona Skill. There is no Work Skill half here; the output is about how a person talks and reacts, not how they do a job.

The celebrity family ships with what the project calls a six-dimension research toolchain. Its stages run from subtitles, to transcript cleanup, to research merge, to a quality check, and the aim is to organize observable decisions, expression and mental models of public figures, creators or fictional characters. This is the family that works from public sources rather than private messages, which makes it the least fraught use of the tool.

The consequence is that the same input produces different profiles depending on the family you choose. Running a mentor's interview transcripts through the relationship family will give you emotional triggers and repair patterns, not their coding standards.

Eight agent hosts read the Skill locally, Grok Bot does not

The earliest version only ran in Claude Code. Distilly now documents native local Skill discovery across eight agent hosts: Claude Code, Hermes, OpenClaw, Codex, DeepSeek Harness, Pi, Grok Build and OpenCode. In each of these, a profile packaged as an Agent Skill can be dropped into the host's local Skill location and picked up without a separate server.

Grok Bot is listed separately as a preview. Grok Bot supports saved or private Skills, but its official documentation does not describe direct local SKILL.md imports, so the project treats it as a saved-Skill workflow rather than a native host. If Grok Bot is your main assistant, you are on a path the project itself marks as a preview, and the drop-in experience available in the other eight hosts does not apply. The repository also links to agentskills.io, which ties the output format to the wider Agent Skill convention rather than a format private to this project.

A Python toolchain behind a Node launcher

GitHub reports Python as the primary language, and the tools/ directory, ruff.toml, requirements files and tests/ folder fit that. The distribution wrapper, though, is a Node package. package.json declares type module, requires Node 18 or newer through its engines field, and ships bin/, SKILL.md, prompts/, tools/, both requirements files, INSTALL.md, INSTALL_EN.md, LICENSE and CITATION.cff. Before packing, a prepack hook checks the package:

node bin/distilly.mjs --check-package

publishConfig points the package at the GitHub Packages registry at npm.pkg.github.com rather than the default npm registry, so the scoped @titanwings/distilly package is published there. In practice you need both runtimes: Node 18 or newer for the launcher and Python for the tools it drives. The step by step route lives in INSTALL.md and its English counterpart INSTALL_EN.md, and the README is translated into Chinese, Spanish, German, Japanese, Russian, Portuguese and Korean under docs/lang/.

The only tagged release is a March demo

The single GitHub release is v0.01, labelled demo, published on 2026-03-30. Since then the project renamed itself, added seven hosts beyond Claude Code, split work into three families and set its package.json version to 1.0.0. The last push to the default branch was on 2026-10-13. None of those later changes sit behind a tag.

The consequence is simple. If you download the release archive, you get the March demo of the colleague-only tool, not Distilly. To get what the README describes, work from the dot-skill branch, and record the commit you installed so that you can reproduce a profile later. There is no tagged changelog to compare against, so the dated update notes in the README are the closest thing to release history: the community gallery opened on 2026-04-07, the roadmap went live on 2026-04-13, the technical report followed on 2026-06-01, the project passed 20K stars on 2026-08-13, and the Distilly rename landed on 2026-08-24.

A profile is only as faithful as the sources behind it

Distilly is careful about its claim. It builds a source-grounded Person Profile from observable experience, decision patterns, expression and ways of working, and it states that it does not claim to clone the person. That is an honest boundary, and it is also the main limit on what you get. A profile built from a few chat exports will reflect those chats; judgment the person never wrote down or said on record cannot appear in it.

The second limit is ethical rather than technical. The relationship family is explicitly pitched at exes, partners, parents and old friends, and the colleague family at people who have left. The tool turns their messages into an agent anyone holding the files can run. Nothing in the pipeline asks whether the person agreed, so that decision rests entirely with you, and for workplace data, with your employer's rules on exporting Lark, DingTalk or Slack history. The community gallery at titanwings.github.io/colleague-skill-site, which grew to 215 skills from 165 contributors according to the June technical report on arXiv, shows how readily these profiles get shared once made.

Editorial conclusion

Distilly suits people who already run one of the eight supported agent hosts and want a reusable profile of a departing colleague's standards and workflows, a public figure's documented reasoning, or their own working style, built from sources they legitimately hold. It is a poor fit if you expect a faithful clone of a person, a stable versioned release, or a Grok Bot setup that loads local SKILL.md files. Before using it, read INSTALL_EN.md on the dot-skill branch rather than the v0.01 demo tag, and decide whether you have the right to feed someone else's messages into a profile at all.

Frequently asked questions

Official sources

1. Official README
2. Project repository
3. Release notes