Prompts & Writing
Getting Consistent Tone Across a Brand When Everyone Uses AI
By Jim Vernon, Editor, AI Intelligence International · Published 16 February 2026 · Reviewed against our editorial standards · About the author
Brand voice used to be maintained by a small number of writers with shared instincts. Distributed, assisted writing removes that mechanism: ten people prompting independently produce ten variations of a generic register.
The fix is to make voice explicit and mechanical enough to be applied by anyone, including a model.
Key takeaways
- Voice as rules, not adjectives: 'Friendly but professional' is unusable.
- Examples with annotations: Collect five short pieces that are unambiguously on-voice and annotate why.
- One system prompt, distributed: Package the rules and examples into a single system prompt and distribute it as the default for anyone writing customer-facing text.
- Review that scales: Move from reviewing every piece to sampling plus a mechanical pre-check.
Voice as rules, not adjectives
'Friendly but professional' is unusable. 'We use contractions. We never use exclamation marks. We say what something costs in the first two sentences. We never call our customers users' is usable.
Write ten to twenty rules of that kind. They should be checkable by someone who has never met your brand.
Include three banned constructions and three signature ones. Contrast teaches faster than description.
Examples with annotations
Collect five short pieces that are unambiguously on-voice and annotate why. Then collect three near-misses and annotate what is wrong.
Near-misses are the more instructive set, because on-voice examples are easy to imitate superficially.
Refresh annually or when positioning changes.
One system prompt, distributed
Package the rules and examples into a single system prompt and distribute it as the default for anyone writing customer-facing text.
Version it, own it, and test it against five real tasks whenever it changes.
Make it available in whichever tools people already use rather than requiring them to visit a document.
Review that scales
Move from reviewing every piece to sampling plus a mechanical pre-check. The pre-check is the rule list; the sample catches judgement problems.
Publish the caught errors internally. Shared examples of drift correct behaviour faster than another training session.
Nominate one editor with final say. Voice by committee converges on the generic register you were trying to avoid.
Channel differences
Voice is constant; register varies by channel. Write the variation down: what changes for social, for support replies, for documentation, for sales email.
Without explicit variation rules, people either apply the marketing voice to a refund apology or abandon the voice entirely.
Two lines per channel is usually enough.
Testing whether it held
Quarterly, take twenty recent published pieces, strip the branding, and ask three colleagues to identify which are yours. Recognition rate is your voice metric.
If recognition is poor, the rules are too weak or unused. If it is strong, you can safely distribute writing more widely, which is the point of the exercise.
Examples beat adjectives, every time
A guide that says 'friendly but professional' produces nothing repeatable, because five people interpret it five ways and a model interprets it as the average of all business writing ever published.
Replace the adjectives with pairs: a sentence written badly and the same sentence written the way you want it. Twenty pairs teach tone more effectively than ten pages of description, and they can be pasted straight into a prompt.
Add a short banned list — words, openings and formulations you never use. Prohibitions are the fastest-acting part of any tone guide because they are unambiguous.
Distribute it where the writing happens
A tone guide in a folder is decoration. The working version is a block of text people paste into prompts, or a saved instruction in the shared assistant that applies automatically.
Keep one canonical copy and link to it rather than copying it into ten places, or you will have ten slightly different tones within a quarter.
When someone joins, give them the pairs and three published pieces you consider correct. That is a faster calibration than any onboarding session.
Audit tone the way you audit anything else
Once a quarter, take ten recent public pieces and mark them against the guide. You are looking for drift in one direction — everything getting slightly more formal, or slightly more breathless — rather than individual lapses.
Drift usually traces to a single source: one heavily-used prompt, one prolific author, or a model default that changed. Fixing the source fixes the next hundred pieces.
Update the guide from what you actually publish and approve. A tone guide that describes an aspiration nobody writes to is a document you will stop consulting.
Record the change in one line with a date, so the next audit can tell deliberate evolution from unnoticed drift. Brands are allowed to change how they sound; they are not helped by changing without deciding to.
Frequently asked questions
Can a model learn our voice from examples alone?
Partly. Examples plus explicit rules performs considerably better than either alone, because rules cover cases the examples do not.
How many rules is too many?
Beyond about twenty, adherence degrades and contradictions creep in. Enforce the remainder through review.
Should support and marketing share a voice?
Share the voice, vary the register, and write the variation down. Customers notice inconsistency between channels quickly.
Who owns brand voice now?
Whoever owns the system prompt and the example set. In practice that should be a named editor, not a committee.
Can one tone cover every channel?
One personality, adjusted per channel. Support replies, documentation and social posts share values but not sentence length or formality.
How many examples are enough?
Around twenty pairs covers most situations. Beyond that, additions tend to repeat rather than clarify.