How do you write tone-of-voice rules AI actually follows?
Most tone-of-voice guides are written for humans — adjectives, mood boards, “be confident but approachable.” A model can't act on that. Rules an AI follows are concrete, testable, and enforced before anyone reviews the draft.
Brand
“A model can follow ‘cut any claim you can't photograph.’ It can't follow ‘be authentic.’”
The answer in 30 seconds
- →Write rules a model can act on: verbs and constraints, not adjectives. “Sentences run 8–20 words” beats “be punchy.”
- →Put the rules in the brand memory so every agent reads them on every task — not pasted into one prompt that's forgotten by the next.
- →Make them testable: a rule you can't check against a draft is a mood, not a rule.
- →Let the rules learn: every edit and rejection at the gate feeds back, so the voice tightens over weeks instead of staying frozen in a PDF.
Why do models ignore normal brand guidelines?
Because most guidelines describe a feeling. “Confident, warm, human” tells a person a lot and a model almost nothing. The rules that hold are mechanical: one idea per sentence; verbs do the work (drafts, checks, flags, signs off) — never enables, streamlines or optimises; hard length variation on purpose. A model can follow “cut any claim you can't photograph.” It can't follow “be authentic.”
What does a rule a model follows look like?
Concrete and checkable. Instead of “keep it punchy,” write “sentences run 8–20 words, with hard variation — some are four words, on purpose.” Instead of “avoid jargon,” give the banned list: no streamline, optimise, leverage. Instead of “be honest,” write “one jagged honest detail per page — a real constraint builds more trust than ten superlatives.” Every rule should be something you could grade a draft against.
[ figure — a reviewer's edit, kept by the brand memory ]
Where do the rules live so agents actually use them?
Not in a prompt. A prompt lives in one chat and dies with the tab. Voice rules belong in the brand memory — read by every agent, on every task, before it drafts. That's the difference between a guideline nobody opens and a rule that shapes the first draft.
How do the rules get better without a rewrite?
The approval gate is the feedback loop. When a reviewer edits “leverage” back to “use,” that correction flows into the brand memory. The voice you get in month three is sharper than week one — not because you rewrote the guide, but because your people corrected the drafts and the system kept the corrections.
Questions your team will ask
Can't I just prompt the model with my voice each time?
You can, once, badly. It won't persist, won't be shared across your team, and won't learn from edits. The brand memory does all three.
What makes a rule “testable”?
You can hold a finished draft against it and get a yes/no. “8–20 word sentences” is testable; “be engaging” isn't.
How many rules is too many?
Enough to be checkable, few enough to be read. One idea per rule; one promise per section.
Do the rules replace my editor?
No. They mean your editor stops fixing basics and starts making the judgment calls only a person can.
Tone of voice, in four words
Plain first, clever never.
Write the rule a model can grade a draft against — and let your team's edits keep sharpening it.