id·i·o·lect

/ˈɪdiəlɛkt/ · noun · linguistics

1. The way of speaking that belongs to exactly one person: their shorthand, their rhythm, their tells.

2. What howitalk, an open source tool, learns about you from your own AI chat history, writes into a file you own, and hands to any AI you talk to.

Existing profiles teach the AI to write like you. This teaches it to read you.

npm install -g howitalk howitalk extract # parse your Claude Code transcripts, locally howitalk distill # write ~/.howitalk/profile.md, a file you own howitalk install-hook # inject it into every session

GitHub npm Format spec Prior art, verified

The problem

You explain. It half-understands. You explain again.

A person sitting next to you gets what you mean on the first try, because they know you. The AI has no such knowledge. You compress six asks into one run-on message and it answers the last one. The usual advice is "write better prompts." That advice is backwards: the fix is not you learning to talk like a machine, it is the machine learning how you already talk.

What it does

A personal interpreter, in a file you own

Learns from behavior: Distills your profile from your real transcripts (Claude Code history, or a ChatGPT export), not from a questionnaire.
A file, not a vault: profile.md is plain text on your machine. Read it, edit it, carry it to any AI. An MCP server and interop exporters make it portable.
Learns overnight, in the open: Opt-in morning refresh studies yesterday's sessions; every change is archived so you can diff what it learned.
Decodes on demand: howitalk decode translates a messy message into an explicit prompt. Shown to you, never silently substituted.
Measures itself: A blinded benchmark ships in the tool: paired responses to your own past prompts, judged without knowing which used the profile.

The honest numbers

We publish our benchmark even though it did not flatter us

Our own blinded test found no measurable effect on single-response quality (42 percent win rate, CI 26 to 61, n=27). Frontier models already parse messy one-shot prompts well. Most tools would bury that number; it is in our README.

The open question is session pace: how often you have to re-explain across real work. The author's correction-rate baseline (3.8 percent of messages) was frozen before the tool went live, and the before and after comparison is running now.

It took three benchmark rounds to get an instrument honest enough to trust. The first two were invalidated by our own audits. Every flaw and fix is in the commit history, and the verified landscape of neighboring tools lives in PRIOR-ART.md, because our first "nobody does this" claim was wrong and an outside reviewer caught it.

Design commitments

Local-first: transcripts, profile, and benchmark data never leave your machine; model calls run through your own claude CLI under your account. Reading, not writing: the profile interprets your input and contains no guidance for imitating your voice. Measured, or it did not happen.