Drafts that sound like you, learned from how you actually talk.
Every AI writing tool tries to copy your voice from your published posts - the most polished, least honest sample of you that exists. Meanwhile, if you use Claude Code, months of your real voice are already sitting on your disk: the transcripts of you talking unguarded about your own work. This skill reads those, builds a private style guide from them, and then turns one sentence into one draft that sounds like you - and every edit you make teaches it.
No API key. No signup. No server. Claude Code does all of it, on your machine, on the subscription you already pay for.
Paste this into Claude Code:
Install the voice skill from https://github.com/rb17080/voice-skill into ~/.claude/skills/ and then run it.
Or by hand: copy skills/voice/ into ~/.claude/skills/ and type /voice.
First run - the mine. It reads your Claude Code transcripts across every
project and builds ~/.voice/voice-source.md: your recurring phrases, your
registers, the constructions you always reject, your best real sentences,
word for word. Takes a while on a big history. Go get coffee.
Every run after - the draft. Give it one sentence about what you want to say. You get one draft back in your voice. Edit it however you like: the difference between what the machine wrote and what you kept is saved, and the next draft is written knowing it. It gets sharper the more you correct it.
Everything lands in ~/.voice/, outside every git repo, and stays on your
machine. That file contains your unguarded words about your own work - treat
it like a diary, because it is one. Nothing is uploaded anywhere by this
skill, ever.
Delete ~/.claude/skills/voice/ and ~/.voice/. That is the whole footprint.