How to build a brand voice guide your AI actually follows

My first brand voice guide said my tone was “direct, warm and occasionally funny”. I pasted it into every AI prompt for about four months and my output sounded exactly like everyone else’s output. It took me embarrassingly long to work out why.

Those three words describe what every language model already thinks it’s doing. Telling a model to be direct and warm is like telling it to write in English. It was already trying.

A voice guide only changes anything when it contains rules that could be broken.

Why adjectives fail

Run the experiment. Ask for a paragraph with no instruction, then ask again with “write in a friendly, conversational, authentic tone”. Compare them. They’ll be near-identical, because the default is already tuned toward friendly and conversational.

Useful instructions are the ones that constrain. “Never use a sentence longer than 25 words.” “Never open with a rhetorical question.” “Use contractions everywhere.” “Never use the word ‘leverage’.” Each of those is testable, and each visibly changes the output.

The rule I’d apply to every line of a voice guide: if the model could ignore it and you couldn’t tell, delete it.

Step 1: Collect your best work

Pick five to ten pieces you’re genuinely pleased with. Same format if possible, since newsletters and long-form articles have different rhythms and averaging them produces mush.

Be honest here. Not the pieces that performed best, the ones that sound most like you. Those aren’t always the same, and the guide is about voice rather than results.

If you don’t have five pieces yet, write one properly with no AI involvement and use that. One authentic sample beats ten approximations.

Step 2: Extract the patterns

This is the part AI is genuinely good at, because it’s analysis rather than generation.

Paste your samples and ask:

Analyse these writing samples and describe the author’s style as a set of mechanical rules another writer could follow. Cover: average and maximum sentence length, paragraph length, use of contractions, how pieces open, how they close, recurring structural patterns, punctuation habits, vocabulary level, use of first and second person, and any words or constructions that appear repeatedly. Be specific and quantitative where possible. Do not use vague adjectives.

What comes back is frequently surprising. Mine told me I open roughly two thirds of my pieces with a first-person admission of a mistake, that I almost never use semicolons, that my paragraphs are unusually short, and that I use “genuinely” far too often. All true, none of it things I’d have said about myself.

That last one is useful in the other direction: a tic you didn’t know you had is a tic you can now decide about.

Step 3: Write the rules

Turn the analysis into a document. The sections that actually earn their place:

Mechanics. Sentence length, paragraph length, contractions, punctuation preferences. Boring and effective.

Structure. How pieces open, how they close, whether you use subheadings as questions or statements, whether you use bullets and when.

Person and stance. First person or not, whether you address the reader as “you”, whether you make claims or hedge them.

Vocabulary level. Do you explain jargon or assume it? What reading level are you aiming at?

The do-not list. The most valuable section, and the one people skip.

Step 4: The do-not list

What you never say defines voice more sharply than what you do.

Mine includes: no “in today’s fast-paced world”, no “let’s dive in”, no rhetorical questions as openers, no exclamation marks, no “game-changer”, no em-dash-heavy sentence construction, and no sentence that begins “But here’s the thing”.

Some of those are personal taste. All of them are things models reach for constantly, which is exactly why banning them makes output stop sounding like AI. The generic register isn’t a tone, it’s a set of specific phrases, and you can name them.

Build this list by reading AI output and noting every phrase that made you wince. It grows over time and it’s the section that improves your results fastest.

Step 5: Test it properly

A voice guide is a hypothesis until you test it.

Take a piece you wrote yourself and had nothing to do with the samples. Now give the model the same brief plus your guide, and compare. The gap between the two tells you what’s missing.

Then iterate. Every time AI output feels off, work out which rule would have prevented it and add that rule. After a month of this the guide is genuinely good, and it got there by accumulating specifics rather than by being clever upfront.

Where to keep it

It needs to be somewhere you’ll actually use it, or you’ll skip it when rushed, which is exactly when generic output slips through.

Tools with a saved brand voice apply it automatically, which removes the discipline problem entirely. Jasper does this well and it’s a substantial part of why I pay for it, covered in my Jasper review and compared against the general models in Jasper vs ChatGPT.

With a general chat model, save it as a custom instruction or keep it in a document you paste. Both work; the difference is whether forgetting is possible.

My take

Spend two hours on this once and every piece of AI-assisted writing you do afterwards is better. It’s the highest-leverage AI setup task available and almost nobody does it, because it looks like admin rather than work.

The specific shift that made mine effective was deleting every adjective and replacing them with rules that could be violated. “Warm and direct” became “use contractions, keep sentences under 25 words, never open with a question, admit the mistake before giving the advice”. Same voice, described in a way a machine can actually follow.

And keep the do-not list growing. It’s the part that stops output sounding like everyone else’s, because the generic AI register is made of specific phrases and you can simply ban them. For the wider writing stack, best AI writing tools for creators who hate the blank page covers the tools this plugs into.