Distilly (formerly Colleague Skill): turning messages and documents into a Person Profile Skill

The problem Distilly addresses: context leaves when a person does

When a colleague resigns, a mentor graduates or a teammate transfers, the artifacts they leave behind (tickets, documents, chat threads) survive, but the reasoning that connected those artifacts does not. Distilly targets that gap. The README frames the project around three situations: a coworker or mentor who has moved on, a relationship you want to preserve the texture of, and a public figure or author you will never meet but whose position on a question you want represented.

The audience is narrower than the slogan suggests. Distilly is not a general summarization tool. It produces a Person Profile, which the README describes as a portable, source-grounded artifact built from observable experience, decision patterns, expression and ways of working. The current release packages each profile as an Agent Skill, so the practical prerequisite is that you already run an agent host that discovers Skills. If you do not, the output has nowhere to live.

How the distilly creator turns material into a Person Profile

The pipeline visible in the repository is linear and file-oriented. You supply source material plus a description of the person; the canonical creator Skill, named distilly and installed in a distilly directory, processes that input into a Person Profile; the profile is then packaged as an Agent Skill that a compatible host can install and invoke.

The repository splits the work into three character families with different collection strategies and analysis dimensions. The colleague family builds a Work Skill plus a Persona from technical standards, workflows, expression and workplace behavior, and the README lists Lark, DingTalk and Slack collection as supported. The relationship family organizes expression patterns, emotional triggers, conflict patterns and repair patterns into a Persona Skill. The celebrity family ships a six-dimension research toolchain that runs subtitles through transcript cleanup, a research merge step and a quality check.

That family split is the most interesting design decision here. A single extraction prompt would flatten the differences between a workplace decision record and a conflict pattern, so the project keeps separate analysis dimensions per family. The cost is that the material you must gather differs by family, and the README does not present a shared schema across the three.

Installing Distilly and running a first real distillation

The repository ships INSTALL.md and INSTALL_EN.md at the top level, and the README links an Install section. The package manifest declares a Node entry point, so the CLI is invoked through Node 18 or newer. The package is published to the GitHub npm registry rather than the public npm registry, which the publishConfig block in package.json confirms.

npm install -g @titanwings/distilly

After install, the binary exposed by package.json is distilly. The manifest also defines a prepack check, which means the package runs its own validation before it is packed.

node bin/distilly.mjs --check-package

The Python side is separate. requirements.txt lists requests as required, with pypinyin, playwright, slack-sdk, python-docx and openpyxl as optional extras for name-to-slug conversion, Lark-compatible browser login, DingTalk scraping, Slack collection and Word or Excel parsing.

pip install -r requirements.txt

The README's usage flow is: source material plus your description goes in, a source-grounded Person Profile comes out, and the profile lands in your agent host. Install the creator Skill into a directory named distilly, then invoke it from the host. What you should see is a generated Skill directory your host can discover locally. The README does not document the exact prompt sequence for each family, so expect to read SKILL.md and the prompts directory before your first run.

Where Distilly is the wrong tool

The README is unusually direct about the central limitation: the profile is built from observable material and, in the project's own wording, does not claim to clone the person behind it. That is a meaningful boundary. If the material contains no instance of a person making a particular kind of decision, the resulting Skill has nothing to ground a response in, and a confident answer on that topic would be unsupported rather than distilled.

The second limitation is host-dependent. The README states that native local Skill discovery is documented for Claude Code, Hermes, OpenClaw, Codex, DeepSeek Harness, Pi, Grok Build and OpenCode. Grok Bot is listed separately as a saved-Skill workflow preview, and the README notes that Grok Bot's official documentation does not describe direct local SKILL.md imports. If your host is not on that list, or if you depend on Grok Bot specifically, the packaging step is not a solved path.

Third, the collectors assume access. Lark, DingTalk and Slack collection depend on the optional packages in requirements.txt, including Playwright for browser login and scraping. That is a heavier dependency than a pure text pipeline, and it means the quality of a colleague profile is bounded by what you can actually export from those platforms.

Distilly compared with a plain prompt or a generic summarizer

The obvious alternative is to paste the same material into a long context window and ask the model to role-play the person. The difference is in what persists. A prompt is per-conversation; Distilly's output is a Person Profile packaged as an Agent Skill, which the README describes as the reusable output and which supported hosts can install and invoke repeatedly. Reuse across sessions and across hosts is the actual product.

The second alternative is a generic document summarizer. Summarization compresses what the material says. Distilly's stated dimensions are experience, decision patterns, expression and ways of working, which means it is trying to extract how a person decides, not just what they wrote. Whether that extraction is better than a careful summary is not something the README quantifies; the project points instead to a technical report on arXiv, cited as 2605.31264, and to a community gallery the README says grew to 215 skills contributed by 165 people.

A third point of comparison is scope. A summarizer handles one document well. Distilly's celebrity family explicitly runs a multi-stage toolchain (subtitles, transcript cleanup, research merge, quality check), which suggests the project expects messy, long-form input and is willing to spend extra steps cleaning it. That is more machinery than most one-off questions justify.

Maintenance, upgrade cost and the MIT licence

The repository is not archived, and the last push was on 2026-03-30. The only release listed is v0.01 (demo), dated the same day. Treat the version number as a signal: this is early, and the README records two naming and scope changes after that release, including the rename from Colleague Skill to Distilly on 2026-08-24 and a roadmap published on 2026-04-13. A rename of the creator Skill means any instructions or scripts that reference the old name need updating; the README says the former name remains for search continuity and project history, so the repository itself still answers to colleague-skill.

Upgrade cost is concentrated in the optional dependencies. requirements.txt pins lower bounds rather than exact versions (requests>=2.28.0, playwright>=1.40.0, slack-sdk>=3.27.0, python-docx>=1.1.0, openpyxl>=3.1.0, pypinyin>=0.48.0), so a fresh install can pull newer releases than the author tested. Playwright in particular needs browser binaries beyond the Python package, and the README does not describe that step.

The licence is MIT, declared in both the LICENSE file and the license field of package.json. MIT is permissive, but it grants rights in the code, not in the source material you feed the pipeline. If you distill a colleague's Slack history or a family member's messages, the consent and privacy question sits entirely with you, and the repository does not address it. Nothing here is legal advice; if the material belongs to an employer or another person, that is worth resolving before you run the collector.

Editorial conclusion

Adopt Distilly if you already run a Skill-capable agent host such as Claude Code, Codex, OpenClaw, Hermes, DeepSeek Harness, Pi, Grok Build or OpenCode, and you have a concrete pile of material (exported Lark or Slack history, transcripts, documents) you want organized into a reusable profile. Do not adopt it if you expect the Skill to reproduce a person's judgment on questions their material never touched, or if you need a documented accuracy guarantee; the README frames the output as source-grounded and explicitly says it does not claim to clone the person behind the material. Before committing, check the install path in INSTALL.md and INSTALL_EN.md, confirm which host you are targeting, and verify that your host actually performs local SKILL.md discovery, since the README notes Grok Bot does not document direct local SKILL.md imports.

Frequently asked questions

Official sources

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