Text that sounds like you.

[Trained on you]

It learns how you write, from what you wrote

A style guide per place you write, counted from every message of yours, with verbatim examples. Then a profile per person you write to a lot.

## Voice
- one thought per message, 7 words median
- lowercase opener (92%), no final full stop (96%)
- asks for a decision at the end: "wdyt?", "ok?"
- never a greeting, never a sign-off
## Examples
- quick one: can we ship the fix before the demo?
- makes sense, let's do the smaller one first. wdyt?

## Pablo
- language: es 95% · en 5%
- length: median 6 words, 88% under 12
- laughter: "jaja" (44) · emoji: 🙏 (18), never at the start
- greetings: none; closings: none
## Examples
- sí sí cuenta conmigo
- te paso la ubicación luego 🙏

## Voice
- opens "Hi <first name>," (78%), closes "Thanks," + name (61%)
- 48 words median; the ask in the first two lines
- lists become dashes, 2 to 4 items
- numbers as digits, dates as "1 March"
## Examples
- Thanks for sending this over. Two things:
- Can we move the start to 1 March?

## Voice
- plain sentences, 14 words median, one idea each
- the decision first, the reasoning after
- "so that" to tie a change to its effect
- sentence-case headings; no exclamation marks
## Examples
- We moved the check to the client so that failures show up before the upload starts.

## Voice
- imperative, 22 words median: what, where, how to check
- paths, flags and identifiers verbatim: tests/refresh.py, -q
- the constraint before the task
## Examples
- Freeze time with the fixture in conftest.py and run the suite with -q.
- Keep the public API; only change the adapter.

Invented guides; yours are counted from your own messages, read straight from the tools, nothing to export. A weekly refresh pulls what is new and relearns a style only when it grew. voz undo <style> puts the previous one back; your ## Overrides survive every relearn.

[Right for where it lands]

Not what was heard. What you would have typed here.

When you press the key, voz reads the active window: which app, which conversation, who you are replying to. The text comes out right for that place.

Under the hood: three speech engines hear at once (a cloud model, local whisper, optionally a second cloud model) and where they disagree is where the misheard word is; the screen is OCR'd while you still talk, never while locked, never a password manager. About three seconds after you let go, the text is at your cursor. If the model is down, the transcript is typed and a notification says so. → How it works, step by step

[Yours]

Nothing comes to us. There is no us.

voz has no backend. What leaves your machine goes to the providers you chose, with your own keys. Everything learned stays in plain files in your home.

[Two ways in]

Your voice, or your agents. Same output.

The writing pass behind the microphone is also a CLI. Anything an agent drafts on your behalf can land in your voice too.

$ voz write --style slack "rough draft of what to say"
$ echo "draft" | voz write --style email --destination "reply to Dana"
$ voz write --style email --screen thread.txt "draft"
# the microphone's route, from a raw transcript:
$ voz process --json <<< "um so the sync is slow"

[Sources]

Where it reads what you wrote

Live sources are read from the tools you already use, incrementally, with your own login. Exports are read once and frozen. One skill per source.

• WhatsApp wacli
• Slack your agent's login
• HEY hey
• Gmail gws
• Google Docs gws
• Coding agents local transcripts
• Exports Slack zip · WhatsApp txt · mbox · Telegram · X · LinkedIn · Markdown

live, refreshed weekly  ·  once

[Roadmap]

Now, next, later

Intentions, not dates. "Now" works today; the rest is the order we mean to build it in.