Every team eventually writes a tone preset. Most of them start with a sentence like this:
“Our tone is warm but professional, conversational but never casual, confident but humble, concise but thorough.”
That sentence is technically a tone preset. It is also exactly the sentence every other tool’s tone preset begins with, which is why nothing it produces ever sounds like the team that wrote it. The vagueness forces the model toward the average of every prose style it has seen. The result is the writing equivalent of elevator music: technically fine, instantly forgettable.
Start with examples, not adjectives
Strip out the adjectives and replace them with two or three before/after pairs pulled from real outgoing work. “Warm but professional” becomes a Slack standup from a senior engineer; “concise but thorough” becomes a two-line release-note entry. Models can pattern-match a concrete shape. They cannot pattern-match adjectives.
If your team opens posts with a one-sentence verdict and unpacks in numbered sections, say that. If they never use an em-dash, say that. If the brand voice says “let’s” instead of “let us” and “ship it” instead of “we will deploy that next week,” spell those uses out by name. The more falsifiable the rule, the less the model has to guess.
Encode rhythm, not only vocabulary
This is the part most presets skip. Two pieces of prose can use the same vocabulary and still feel completely different. Compare:
“We pushed the fix to staging yesterday and are watching metrics this morning.”
versus:
“Yesterday’s push to staging is holding. We’re watching the metrics and will report back at end-of-day.”
Same words, different tempo. Short sentences, the occasional fragment, a dash used to interrupt, and one verb-tense rule (present-tense for the team, past-tense for the system) usually carry more of your voice than another adjective does.
A small exercise: take the last ten outgoing messages from your team and mark them up for sentence length. Most teams you would guess at a 60 / 20 / 20 split — 60% short, 20% medium, 20% longer and broken up with a list or a dash. Most “generic” presets output something closer to 30 / 40 / 30. The first split will feel like your team. The second will not.
Show what “off-brand” looks like
A preset with no counter-example teaches the model the centroid. Every “generic AI tone” we recognise came from a centroid. Include one or two snippets explicitly labelled “do not write like this” — the over-hedged apology opener, the sign-off no human on your team has ever actually used (“Please don’t hesitate to reach out if you have any further questions”). Models are bad at negation in natural language, so positive framing of what to write is more reliable than a list of forbidden phrasings. But one or two pointed counter-examples still help the model locate the centroid by marking its edges.
A workable first draft
If you only have an hour, do this: take three real outgoing messages from the past two weeks. Find two patterns that repeat — one about vocabulary (“we” not “the team,” “ship” not “deliver,” contractions by default) and one about shape (the sentence-length mix, how often you reach for a list versus prose, how often you sign off with a one-line next step). Write each pattern as one short paragraph, then use them.
You will sound like your team far more often than the brand-voice document does. The voice document is a wish. The two patterns are a measurement.