distilly for Claude Code and Codex: turn a person's judgment into a skill

distilly takes a person’s documents and conversations and turns their experience, judgment, and voice into a skill an agent can use. The repository by titanwings has collected 24,822 stars, and it works in both Claude Code and Codex. This post covers what problem it solves, who should install it, how the first run goes, and the check to make before you feed it anything.

What problem it solves

When a team lead leaves, the team keeps their documents but loses the reasoning behind them. The design documents say what was decided, and the review comments say what was rejected, but nobody can ask the person why anymore. distilly reads that material and packages the patterns it finds into a profile skill named {character}-{slug}. Afterward, an agent can answer the way that person would have answered, grounded in what they actually wrote.

The material can come from many places. The README covers Lark, DingTalk, Slack, WeChat, PDFs, screenshots, and emails. When the profile gets a judgment wrong, you point it out in conversation, and distilly records the correction in a separate layer that you can roll back. New material is analyzed incrementally and merged into the relevant sections without overwriting earlier conclusions.

Who it fits

The chapter rates the fit for four situations. Three of them are enough to decide.

Situation	Fit	Reason
Handover from a team lead	Good fit	Preserves how they decide
Organizing a senior colleague’s reviews	Good fit	Uses long-form material
A business contact with no source material	Poor fit	Source quality determines the result

The README recommends design documents and review comments written by the person as the best input. Conversation logs that are mostly small talk carry lower priority, so a thick folder of chat history is not a substitute for a few real decisions written down.

Install and first run

Clone the repository into the skills folder for your agent.

# Install globally for the Claude Code user
git clone https://github.com/titanwings/distilly ~/.claude/skills/distilly
# Install globally for the Codex user
git clone https://github.com/titanwings/distilly ~/.agents/skills/distilly

Then ask the agent to create a profile and hand it the source material.

Create a profile of [person]
[documents or conversations from the person]

The run starts by choosing a type: colleague, relationship, or celebrity. You then provide a description, material, or both, and the README says both are optional. When the profile is ready, you install the generated skill by naming a host in the installation tool. The README lists eight hosts with local skill discovery, including Claude Code and Codex.

What trips people up

Consent comes before collection. The README says to check policies and consent before you collect anyone’s conversations and documents, and never to commit tokens to the repository. A profile built from a colleague’s messages is still their messages, so settle that question with them and with your company before you run the collectors. The collectors also carry setup cost. Automatic collection needs Python dependencies and Playwright, Slack needs an administrator to install the app, and Slack’s free plan limits message history to 90 days.

Where this comes from

This post adapts part of chapter 01-03 of The Claude Code & Codex Skills Guidebook. The full chapter covers the celebrity pipeline that organizes subtitles and transcripts before six strands of research, and the full list of what the repository includes, from the three profile types to the source collectors, and the book has 52 more skills like this one. Read the full chapter in The Claude Code & Codex Skills Guidebook on Leanpub.