First, your posts
At sign-up you give Sonar your LinkedIn profile or connect X, and your published posts are imported. They become the voice section of your Intelligence Pack: the patterns in how you open, argue and close; a banned list of the tics you never use; example sentences pulled from your posts verbatim; and a rubric of checkable rules a writer could follow to sound like you.
Before the model writes any of that, Sonar measures the posts. How many words your sentences average. What share run under eight words, and what share over twenty. How many sentences a paragraph carries, and how often a paragraph is a single line. How many sentences in a hundred are questions. How many emoji a post carries. Those numbers are written into the profile as measured, with the instruction not to exaggerate the punchiness beyond them, and every draft is held to them.
LinkedIn and X get separate registers from the same identity: where you have real posts on the other platform too, its register is learned from those, so a long-form draft never inherits tweet compression.
Then, your edits
Every draft you edit before posting is compared with the version Sonar wrote. Once a platform has three edited-then-published drafts, a weekly pass reads the differences and distils them into rules with evidence attached: what you keep changing, stated as a directive, with the count that supports it. Those rules ride into every subsequent draft, and you can read and mute any of them in Settings under Voice, in the panel called Learned from you. Three is the floor, so one stray edit never becomes a habit.
You can also just say it. Tell the chat you never want a question as an opener, and it offers a chip: save as a standing voice rule? A yes writes the rule; silence keeps it to that conversation. A stated preference needs no edit evidence, so it applies from the next draft.
And, if you choose, how you speak
Your posts are how you write. How you talk in a meeting is often better material, and the sources that carry your own words can teach it: Meet and Zoom transcripts with their speaker labels, Fireflies transcripts when it is connected with a pasted key, the Slack conversations you picked with their author names, the Notion pages you created yourself, and, again with a pasted key, the lines Granola attributes to you. This is off by default. The switch is "Learn voice from your sources" in Settings under Voice, and the sentence beside it names exactly which of your connected sources it would read.
With it on, once a week Sonar reads those sources for one purpose: the text attributed to you, and no one else's, is distilled into four to six notes on your spoken register. Each note carries a count as evidence and never a quote longer than six words. The notes replace the previous week's, and your posts stay the anchor of your voice. One kind of spoken material needs no switch: the story answers you dictate into Sonar's own interview, in Settings under Intelligence, Your story, are words you gave it directly, and they teach the spoken register on their own.
What it never does
No model is trained on your posts or your words. The voice section is text, sent with each drafting request; under the AI provider's API terms your inputs and outputs are not used to train its models, and the privacy policy says so. A document you upload can change what a draft says and never how it sounds; the voice section wins every argument about style. And the anti-cliche pass that runs on every draft reads your voice profile first: if you write in lowercase, skip full stops or lean on brackets, that is your voice, and it is left exactly as you write it.