Voice, learned from the work itself

Every team has a way it writes. Sentence rhythm, what gets the semicolon, which words are out, which words are signatures. Most teams write it down once in a style guide and then nobody reads it.

This ability learns the voice from the work itself.

What it does

Reads the team’s published pieces and builds a voice profile: sentence-length distribution, vocabulary, structural habits, the words and phrases the team avoids.
Applies the profile to every draft the brain generates so output reads like the team wrote it.
Maintains separate profiles per writer or per surface if the team needs them (longform vs. newsletter vs. board memo).

The reveal

A new staff writer turns in their first piece. The editor runs the voice-match check and gets a list: “three sentences use ‘utilize’ where the team has historically used ‘use’; the lede is twice the average length for this section; the kicker matches your usual pattern.” The edit gets faster, and the new writer learns the voice by reading the diff.

What it doesn’t do

No homogenization. The profile is a reference, not a hard constraint. Writers can write against it on purpose. It surfaces where the draft drifts from the team’s voice; the writer decides whether that drift is the point.

Configuration

Corpus: which published pieces define the voice.
Profiles: one team-wide, or one per writer / per surface.
Strictness: flag every drift, or only the ones above a threshold.

Triggers