Core concepts

Voices

Voice profiles learned from 2–20 samples: what they measure and how to build a good one.

A voice is a reusable profile of how one person writes, learned from 2–20 of their own pieces. Pass its id as voice_id on a rewrite and the output follows that person's habits instead of a generic “professional” register.

What a voice measures

We measure habits a careful reader notices, not topics. The traits object exposes the main ones:

Trait	Example value	What it captures
`sentence_length_sd`	9.4	How much sentence length varies. People vary a lot; models tend to be uniform.
`uses_em_dash`	false	Punctuation habits, including ones the writer never uses.
`numbered_steps`	true	Whether instructions become lists or stay as prose.
`endings`	status_update	How pieces close: a question, a call to act, a plain status update.

The profile also stores a baseline: how the Detect API scores the writer's own samples. That's what makes calibrated detection possible.

{
  "id": "voice_8f2k1x",
  "object": "voice",
  "name": "Founder voice",
  "samples": 2,
  "words": 1197,
  "traits": {
    "sentence_length_sd": 9.4,
    "uses_em_dash": false,
    "numbered_steps": true,
    "endings": "status_update"
  },
  "baseline": {
    "ai_likelihood_mean": 0.19
  }
}

Choosing good samples

Same person, unedited. Avoid pieces that went through heavy editing or an AI tool; you'd be teaching the voice the editor's habits.
At least 1,200 words in total. More helps up to about 10,000 words, then returns flatten.
Match where you'll publish. Five LinkedIn posts teach LinkedIn habits. Use platforms to say where the samples came from.
Vary the topics. Different subjects let us separate style from subject matter.

Samples are for style only

Facts in your samples (names, figures, links) are never copied into a rewrite. The fidelity checker only allows facts that appear in the source you're rewriting.

Using a voice

Pass voice_id on POST /rewrite. Without it, rewrites use a neutral house voice: plain, specific and free of stock phrasing, but nobody's in particular.

You can also pass a voice to POST /detect as calibrate_to, which scores text relative to that writer's baseline.

Limits and lifecycle

Voices aren't billed. Basic API includes 3 voices, Pro API 15, Ultra API unlimited.
Your samples are used to build the profile and then deleted; they are never stored with the voice. The profile keeps style measurements, a detector baseline and a style card that quotes 5–8 short sentences from your samples.
Voices are immutable. To update one, create a new voice with the extra samples and switch ids; old ids keep working until you delete them.
Deleting a voice removes its derived profile (and cancels a build that hasn't started). See Security & data retention.