If you have ever handed a writer or an AI tool the instruction "sound confident and friendly" and gotten back something that sounds like everyone else, the problem usually isn't the writer or the model. It's the instruction. Adjectives like confident, friendly, and human describe a feeling, not a behavior, and neither a person nor an AI writer can execute a feeling. This guide walks you through how to document brand voice guidelines as a set of observable rules, with examples, so both a new hire and an AI writer can follow them the same way. By the end you'll have a repeatable method for writing a brand voice and tone guide, and voice guidelines an AI writer can follow without guessing at what you actually meant.
This article covers the guide itself: the document you write once and hand to anyone drafting for your brand. It doesn't cover turning that document into machine-readable config, the day-to-day habits that keep a team's writing on-brand once the guide exists, or how to review a finished draft against it. Those are separate problems worth their own process. Here, the goal is a usable, specific brand tone documentation you can point a person or a model at and get consistent results.
Define what the voice needs to accomplish
Before you pick a single adjective, decide what the voice is for. Voice exists to build trust, make the brand recognizable, and get readers to understand and act. If you skip this step, you'll end up choosing words that sound nice on a slide but don't change a single sentence of actual copy.
Ask yourself what a reader should understand after reading, what they should feel about the brand, what action they should be able to take, and what would make the brand sound credible or evasive to that specific audience. Then keep three things separate in your document: voice is the consistent personality that stays the same across everything you publish, tone is how that voice shifts for a specific situation, and audience is who you're actually writing for, including what they already know and what they're trying to get done.
You'll know this step is done when your document has a clear audience description, a stated communication purpose, and voice and tone explained as two different things, not lumped together as one vague personality.
Common mistake: starting with a list like "we are friendly, innovative, bold, and authentic." Those words don't tell anyone what to write. Every one of them needs to become a behavior before it's useful, which is exactly what the next step does.
Collect and annotate your strongest examples
Go find the writing that already sounds like you at your best: a welcome email that got replies, a product page that converts, a social post that got shared for the right reasons. These become your source material, but only if you annotate them.
Don't just paste an example into the guide and call it done. Write next to it what makes it work. Does it open with a direct answer or a question? How long are the sentences? Does it use contractions? Does the writer talk straight to the reader? How does the copy move from the reader's problem to an explanation to what to do next? Which specific words make it recognizably yours and not something a competitor could have written?
Mark each example clearly as on-brand, off-brand, or situational. An unmarked example is worse than no example, because a writer or a model reading the document later has no way to know whether an older piece of copy is something to copy or something to avoid. The same idea applies when you train an AI tool on your own best writing: an annotated example teaches faster than an adjective ever will.
You'll know this step is done when a new writer can look at an example and explain which specific behaviors to repeat, not just that the example is good.
Common mistake: picking examples only because they performed well. A high-performing post might carry a campaign-specific tone or a one-off joke that shouldn't become a standing rule. Performance is a signal, not proof that a piece represents your everyday voice.
Turn your adjectives into three to five behavioral principles
This is the step that makes the rest of the document usable. Take the handful of words you'd use to describe your brand and rewrite each one as something a person can actually do in a sentence.
Take "confident" as an example. On its own, it tells a writer nothing. Turned into a principle, it looks like this: open with a direct statement that answers the reader's question, state the outcome before the method, use active voice, and don't weaken a claim you can support with phrases like "we believe" or "it may be possible." Compare a vague, hedged sentence like "in today's fast-changing environment, it may be possible for teams to improve their process by considering a number of different approaches" against a direct one: "start by documenting the decisions your team makes repeatedly." The second version is direct without being aggressive, and that distinction matters. Confidence should never slide into hype or claims the evidence doesn't support.
Write three to five of these principles, no more. Each one needs a one-sentence definition, three to five things a writer should actually do, a short list of what the principle does not mean, at least one before-and-after rewrite, and the words you want used and avoided. A good test: hand someone a paragraph and see if they can tell you whether it follows the rule. If they can't, the rule isn't specific enough yet.
Pro tip: write the before version first, using the vaguest, most generic phrasing you can imagine for that trait. It's much easier to fix a bad sentence into a good one than to invent a good sentence from nothing, and the contrast is what makes the rule stick for a reader.
You'll know this step is done when every principle has instructions specific enough to apply to a real sentence and a clear boundary explaining what it doesn't mean.
Common mistake: swapping one adjective for its synonyms. "Bold, powerful, strong, authoritative" is still just a list of words. None of it tells a writer what to type.
Document how tone shifts by situation
Your voice should stay recognizable everywhere, but the way it shows up changes with the moment. An educational post can carry more explanation and a slower pace. A response to a frustrated customer needs more empathy and a shorter path to a next step. A sensitive topic calls for more precision and less humor. A launch post can carry more energy. What stays constant across all of these is your vocabulary, your evidence standards, and your respect for the reader.
Build a short table that names the common situations your team writes for and, for each one, what changes and what stays the same. This keeps the document from turning tone into a totally different personality for every channel, which just gives your writers and your AI tool contradictory instructions to follow.
You'll know this step is done when a writer can shift the emotional temperature or level of detail for a specific situation without changing the brand's underlying identity.
Common mistake: treating every channel as its own separate voice. That produces an inconsistent reader experience, and it's exactly the kind of contradiction that makes an AI writer's output swing wildly between pieces.
Set your word-choice, sentence, and mechanics rules
Word choice does a lot of quiet work. Whether you use contractions, how long your sentences run, whether you write in first or second person, all of it shapes how a reader experiences the brand, even when they can't name why.
Write down your preferred terms for your own products, audience, and outcomes, and the discouraged terms that are technically fine but sound too vague or too formal. List banned phrases, the ones that keep sneaking into drafts and pulling them off-brand, like empty transitions or unsupported superlatives. Note any terminology that has to stay exact for legal, product, or accessibility reasons. State plainly whether the brand uses active or passive voice, first, second, or third person, and how you address the reader.
Mailchimp's public style guide is a useful compact model here: active voice, plain English, and positive language over negative phrasing. You don't need to copy their choices, but the shape of that guidance, short, concrete, and specific to real word choices, is what makes a rule usable.
You'll know this step is done when two different writers, faced with the same sentence to write, would land on nearly identical word choices.
Common mistake: calling your voice "conversational" and leaving contractions, slang, sentence length, and how you address the reader completely undefined. That word alone doesn't tell anyone what conversational means for your brand specifically.
Organize the rules by content element so an AI writer can use them
This is the step where you write brand voice guidelines for AI specifically, not just for the humans on your team. A model performs much better with instructions tied to a specific part of the content than with one long paragraph titled something like "sound like us." Everything from here on is what separates a document a person can skim and internalize from the kind of voice guidelines AI writer can follow line by line without a human filling in the gaps.
Break your rules out by element: openings, headings, claims and proof, explanations, transitions, calls to action, and endings. For openings, tell the writer to start with a direct statement that answers the reader's question, and to skip rhetorical questions or generic scene-setting. For claims, separate fact from opinion from estimate, require a source and a date for any performance claim, and state plainly that nothing gets invented: no made-up statistics, customer results, quotes, or product capabilities. For transitions, allow one only when it shows a real relationship between two ideas, not as a filler phrase. For calls to action, tell the reader what to do and why it matters, without manufactured urgency. For endings, close with the next useful action instead of a repeated summary, and never introduce a new claim in the last sentence.
If you're briefing an AI writer directly, a clear pattern helps: state who the audience is, what the content needs to accomplish for them, which voice principles apply, the element-specific rules above, the preferred and avoided language, and an explicit instruction not to invent facts, sources, or product details, and to flag missing information instead of guessing. Model providers describe the same pattern in their own prompt engineering guidance: specific, structured instructions beat a vague brief. You're not writing one giant prompt. You're making the documentation precise enough that anyone on your team can translate it into whatever workflow they use.
This is also where stored brand context starts to earn its keep. DeepSmith's Deep IQ module holds your brand voice, product facts, and persona details as structured context that every generated draft pulls from, so a writer isn't re-briefing the same rules for every single article. The tool only works as well as the document behind it, though. Storing a vague voice description doesn't make it executable. The specificity has to exist first.
You'll know this step is done when the document tells a writer or a model what to do at each major content element, with an example of acceptable and unacceptable output for at least a few of them.
Common mistake: handing a model a long list of personality words and expecting it to guess at the structure, the evidence standards, and the boundaries on its own. It won't guess right, and it won't guess the same way twice.
Test the guide against real writing tasks and keep it current
A voice guide can read as thorough and still be too vague to produce consistent copy in practice. The only way to find out is to test it.
Pick a handful of real tasks your team actually writes: not a hypothetical exercise, but the formats and situations you cover every week. Run them through your process, whether that's a human writer, an AI tool, or both, and compare the output against your strongest existing examples side by side. Score the results against something concrete: voice match, factual accuracy, clarity, structure, and audience fit are five dimensions that work well together. When something misses, write down the specific rule that should have caught it, not just a note that it "felt off." Then go back and fix that rule, that example, or that word list.
For anything factual, customer-facing, legal, or otherwise high-risk, keep a human review step in place regardless of how well the guide performs. A documented voice reduces ambiguity. It doesn't verify a fact or catch a legal problem on its own. Search engines judge writing by its quality, not by who or what wrote it, which is one more reason a documented voice pays off even outside your own team. Federal guidance on generative AI even recommends documenting a system's limits and how its output is meant to be overseen, which is exactly what a tested, current voice guide does for your own writing process.
Keep a short log of what you test and what changed as a result. When a new hire or a new tool starts producing off-brand copy six months from now, that log tells you whether the guide has a real gap or whether this is a one-off mistake worth a quick note rather than a rewrite. It also keeps the guide honest: a brand voice and tone guide that never changes after its first draft is usually one nobody is actually testing against real work.
You'll know this step is done when you've run the guide against real tasks, traced the failures back to specific gaps in the document, and know roughly how often you need to revisit it.
Common mistake: treating the first version as permanent. Your product changes, your audience changes, and the tools you write with change. A guide that worked well a year ago can drift out of calibration without anyone noticing until the copy starts to feel generic again.
What to do next
Start with the principles, not the whole document. Pick your three to five behavioral rules, write one before-and-after example for each, and test them against a single upcoming piece before you try to document tone variations or a full word list. A guide that's half-finished but specific beats one that's complete but still full of adjectives.
If you're documenting this for the first time, resist the urge to write the whole thing in one sitting. A brand tone documentation project that tries to cover every channel, every content type, and every edge case on day one usually stalls before it's finished. Ship the three to five principles this week, test them against real drafts next week, and add the tone-by-situation table and the word-choice rules once the core behaviors are holding.
Once you have a working draft, keeping it applied at volume is its own separate problem: it's the difference between a document that exists and a voice that actually holds across every piece your team publishes, human or AI-written. If you want to see what it looks like when that stored context feeds directly into production instead of sitting in a shared drive nobody opens, DeepSmith's free trial lets you set up your own brand voice and watch it show up in a real draft.
