Claude Code Self Distill

**Language:** **English** | [简体中文](README.zh-CN.md)

self-distill

[](LICENSE) 

Your conversation habits with AI are uniquely yours. Other people's skill templates never quite fit, but you don't know where to start writing your own. Self-distill helps you find answers from conversations, but you take the lead.

A meta-tool that discovers, distills, and generates personalized skills from your own conversation history.

The routine is simple: run `/self-distill record` at the end of a useful conversation (If the experience is not satisfactory, add `--negative`.), then periodically run `/self-distill extract` in a fresh session. That's it — a **user-driven** closed loop that stays out of your way.

---

Why this tool exists

This tool started as a solution to my own problem.

After using Claude Code for a while, I noticed I kept doing the same things: re-explaining the same constraints, correcting the same directions, adding the same context. All of this could have been captured in a skill once and for all — but the problem was, I didn't know what to write.

I tried organizing things manually, but it was too subjective. I'd include things I *thought* were useful while missing the patterns that actually recurred.

There are similar tools out there: Hermes skill-factory can fully automatically analyze and generate skills; Claude Code's built-in `/insights` generates an HTML report with friction points and behavioral metrics. But both share the same problem — **no filtering mechanism**. Every conversation enters analysis indiscriminately, and noise from low-quality sessions dilutes the real signal, leading to skills that miss the mark.

So self-distill's core principle is: **semi-automatic, with you in control.**

You decide which sessions are worth archiving. The AI finds the patterns in those archives.

---