Measure the voice with numbers, then test it by covering the name

Measure, don't describe

'Median sentence length seventeen words, one sentence in eight under six' is usable. 'Direct but warm' is not, because nobody downstream can check it against anything. Write the specification as a table of measures, not adjectives: sentence length (median and longest, as numbers), signature terms (counted per thousand words, not guessed), words it never uses (a written list), opening and closing moves (named: example-first or conclusion-first), and what it does when uncertain (named: hedges, asks outright, states flatly).

The covered-names test

Strip the names and signature phrases out of a few paragraphs written from the specification and hand them to somebody who knows the person. On one documented test, the reader picked the paragraph written from the properly measured specification over the alternative. That's the test a voice spec has to survive before it's trusted with real drafting: not whether it sounds plausible, but whether it beats a blind comparison.

Write the never-touches list in the same file: every word and construction this voice doesn't reach for. A junior left without one defaults to generic habits, stacking adjectives, reaching for exclamation marks, repeating 'not X, but Y' as if it were insight rather than a tic. A short, specific list beats a long, general one; a junior checking its own draft against ten vague principles will find a way to pass all ten anyway.

What done looks like

A written specification exists with numbers, not adjectives, on every measure, plus a never-touches list of specific habits this voice avoids. The specification has passed its own covered-names test: a reader who knows the person picks the real paragraph out of a blind lineup.