Brand Voice With AI: Style Guides, Few-Shot Examples, and Killing the AI-Slop Tell

Here's the situation almost every marketer using AI has run into. You open ChatGPT, or Claude, or Gemini, doesn't matter which, and you type something reasonable. Write me a launch email for our new feature. Make it professional but friendly. And back comes this email. It's grammatical. It's organized. It hits the points. And it is completely, hopelessly generic. It sounds exactly like the email every other company on earth would get back from the same prompt. It sounds like AI.

That's the problem we're solving today. Not how to write a better prompt in the mechanical sense. You already know how to write a structured prompt. Role, task, context, format. We're not re-deriving that. Today is narrower and, honestly, more valuable. Today is about voice. Specifically, how to make the AI sound like your company and not like the statistical center of the internet.

Let me name the actual skill, because naming it changes how you approach it. The skill is voice transfer. You're taking a model that has one default writing style, and you're pulling it off that default, toward yours, and keeping it there. And here's the thing people miss. That default AI voice? It's not the absence of a voice. It's a voice. A specific one. It's bland, it hedges everything, it's relentlessly upbeat in a corporate-LinkedIn way. The model didn't choose it because it's good. It landed there because it's the average. It's the most statistically likely way to write when you don't tell it otherwise. So you're not adding personality to a blank page. You're overriding a strong, pre-existing personality. That distinction matters, and it's why this takes a little muscle.

Now here's the reframe that I want to stick with you for the whole episode. You are not writing prompts today. You're building an asset. A reusable brand-voice profile. A document. And then you ground the model in that document every single time. The prompt is just the delivery truck. The asset is the cargo. The value lives in the cargo. Once you've built it, you can hand it to any tool in a couple of minutes and get on-brand output. That's why we're going to spend most of our time on the asset, not on any one chatbot.

And one quick foreshadow, half a sentence, I'm not going to elaborate. This voice profile you hand-build today is the exact same artifact that, later in the series, becomes a grounded knowledge base and an agent's standing brand context. Hold that thought. Today it's craft. At the desk. By hand.

The good news under all of this: the model already can write in basically any voice. Hemingway. BuzzFeed. A legal brief. A Slack message to your work bestie. It contains all of those. The default voice is just where the gravity pulls it when you stay quiet. Your job is to fight the gravity. Let's learn how.

Okay. The first thing we have to untangle is voice versus tone, because marketers mix these two up constantly, and the mix-up is the root of why outputs drift on you.

Voice is the constant. It's your brand's personality in writing, and it stays the same day to day. Tone is the variable. It's how that voice flexes depending on the situation and how the reader is feeling. Same brand, same voice, but the tone of a celebration is different from the tone of an apology.

The cleanest articulation of this comes from Mailchimp. Their content style guide is public, it's been a reference for years, and the canonical line is basically this: our voice doesn't change much from day to day, but our tone changes all the time. Think about what that means in practice. When they send a product-launch email, the tone is celebratory. When they have to tell you there was a billing error, the tone is careful and plain. Different tone. Same Mailchimp underneath.

And it's worth knowing what that Mailchimp voice actually is, because they spelled it out, and it's a real, public model you can study. They describe themselves as plainspoken and genuine. They translate technical stuff into human language. And they've got this dry, offbeat sense of humor. They literally wrote that they prefer the subtle over the noisy, the wry over the farcical. And my favorite line in the whole guide: forced humor is worse than no humor. When in doubt, keep a straight face. That's a rule. That's enforceable. You could hand that to a writer or to a model and it would actually change the output.

They've also got four writing goals, and they're refreshingly boring, which is the point. Clear. Useful. Friendly. Appropriate. And underneath those, two more lines worth tattooing somewhere. Write like a human. And, clarity is more important than entertainment.

Nielsen Norman Group, the user-experience research group, has a nice framing for tone specifically. They say tone of voice is the way we tell our users how we feel about our message. I like that because it makes tone emotional and contextual, which it is.

So why does any of this matter for working with AI? Because here's the mistake. A marketer wants to capture their brand, so they tell the AI, be professional but friendly. But professional but friendly is a tone instruction. It's about the emotional register of one message. It is not voice. It doesn't capture the durable, day-to-day personality of the brand at all. So the AI fills in the gaps with its own defaults, and the output drifts, and you can't figure out why. The fix is to separate the two. Your durable asset, the profile, captures voice. Your per-task instruction adjusts tone. Voice in the document. Tone in the moment.

Let's make tone concrete, because Nielsen Norman Group did some genuinely useful work here that gives us something we can hand straight to a model.

They published a piece called The Four Dimensions of Tone of Voice, by Kate Moran. It first came out back in 2016 and they updated it in 2023, so it's been pressure-tested. And the four dimensions are four sliding scales. Not on-off switches. Scales. You place your brand somewhere along each one.

Scale one: funny versus serious. Scale two: formal versus casual. Scale three: respectful versus irreverent. Scale four: enthusiastic versus matter-of-fact. That's it. Four sliders.

A couple of nuances Moran points out. That irreverent dimension? When brands use it well, the irreverence is usually aimed at the subject matter, as a way to stand apart from competitors, not aimed at the reader. You're being cheeky about the topic, not rude to the person. Keep that straight or you'll alienate people.

And there's actual research behind this. Moran ran a study with around fifty participants, showing them the same content written in different tones. The tone variations produced statistically significant differences in how people perceived the brand. The effect sizes were modest, roughly half a point to a point on a five-point scale, but they were real and measurable. A related Nielsen Norman study found that casual, conversational, moderately enthusiastic tones tended to perform best overall, though, and this is important, those qualities don't have to be bundled together. You can be casual without being especially enthusiastic. You mix and match.

Here's a worked example so you can hear it. Take a plain system message: an error has occurred. Now flex the tone. Formal and serious version: we apologize, but we are experiencing a problem. Now casual and enthusiastic: oops, we're sorry, but we're experiencing a problem on our end. Same information. Same underlying message. Completely different feel. That's tone modulation in three sentences.

Now, why are these four dimensions gold specifically for working with AI? Because they turn vague vibes into coordinates. Instead of telling the model professional yet approachable, which means nothing, you can say: funny, three out of ten. Formal, four out of ten. Respectful, eight out of ten. Enthusiastic, six out of ten. That is something a model can actually act on. Numbers on known axes beat fuzzy adjectives every time. You've given it a target instead of a mood.

And that brings us to the heart of this whole episode. The single most important idea. Examples beat adjectives. If you take one thing away today, take this.

Adjectives are lossy. They lose information. When you tell the model be conversational and bold, that phrase could mean a hundred different things, and the model has no idea which one you mean. So it picks its own interpretation, filtered through its defaults. But three paragraphs of your actual best writing? That means exactly one thing. This. Write like this. Examples are nearly lossless. They carry the information that adjectives throw away.

Here's the mechanism, because understanding it makes you better at this. When you give the model an adjective, it interprets that adjective through its own priors, its own defaults. So you say witty, and you get the model's idea of witty. Which is puns. And buckle up. And forced exclamation points. Generic witty. But when you show the model your wit, three real examples of your brand being funny, it copies your register. Your specific flavor. It stops guessing.

The technical name for showing examples is few-shot prompting. Let me gloss that. Zero-shot is when you just give an instruction and no examples. Do this thing. Few-shot is when you give the model a handful of worked examples inside the prompt, so it imitates the pattern. For voice work, the examples are three to five real samples of your writing. Or, even better, and we'll come back to this, generic-to-on-brand rewrite pairs.

That better version has a name too. Contrastive few-shot. Paired examples. Here's why it's the most powerful variant. Instead of only showing the model the destination, your good writing, you show it the before and the after, side by side. Here's the generic version. Here's our on-brand rewrite. Now the model doesn't just learn where to end up. It learns the delta. The transformation. What to strip out and what to add. That's a much richer lesson than just showing the finish line. You're teaching the move, not just the result.

How many examples? Rule of thumb: three to five. That's the sweet spot. Fewer than three and the model doesn't have enough to see a pattern, it's just guessing from one or two data points. More than about five and you're burning up the context window, the model's working memory, without getting proportionally better results. Three to five. That's your range.

And here's a finding worth knowing. Rules plus examples beat either one alone. If you combine a written voice spec with actual examples, style adherence improves over rules-only and over examples-only. This is directional, not gospel, but it lines up with intuition. Tell it the rules and show it the rules. Both. So the profile we build is going to have a written spec and example samples. Belt and suspenders.

Now let's talk about what makes a good example set versus a bad one, because this is where people quietly sabotage themselves.

A good example set is representative, not exhaustive. You don't need every piece you've ever written. You need a diverse handful of slices. A simple case. A harder case. An edge case. Or different formats. The examples should actually be good, because, and I cannot stress this enough, the model copies mediocrity just as faithfully as it copies excellence. And the examples should be yours. Real copy you're genuinely proud of. Not aspirational copy from some brand you admire. Because if you feed it a competitor's writing that you wish you sounded like, congratulations, you just taught the model their voice, not yours.

A bad example set does the opposite. Inconsistent samples, where the pieces don't sound like each other, so the model averages them into mush. Off-topic samples. Samples that contain the exact tics you're trying to kill, which just trains the tics deeper. And committee-written copy with no spine, the stuff that went through six rounds of edits until every interesting edge got sanded off.

Two traps worth naming right here. First, the average-of-your-samples trap. If you mix inconsistent samples, the model literally averages them, and you get bland. Curate ruthlessly. Second, faithful mediocrity. Garbage in, fluent garbage out. The model will reproduce your weak writing with total confidence. So pick your best.

Okay. Now the keystone technique. This is the move that makes everything else easier, and it's almost magic the first time you do it. Reverse-engineering. You let the AI extract the voice profile from your own samples.

Here's the procedure. You paste in three to five of your best, most representative pieces, and you ask the model to analyze the patterns and write you a reusable style and voice profile that you can feed back in later. You're using the model to study your writing and hand you back a spec. It's genuinely good at this. Pattern-matching across text is exactly what these things do.

Let's talk volume, because more real material means sharper extraction. At an absolute minimum, give it a few hundred words. Five hundred plus is better. And if you want a really solid profile, aim for something like three thousand to five thousand words of sample text total. The more genuine writing you give it, the sharper the extraction. It's working from evidence, so give it evidence.

Now, what do you actually ask it to extract? You want measurable, enforceable rules, not vibes. You're turning feel into a spec. So ask it for things like: your typical sentence length and rhythm. Your vocabulary level and register, plain versus technical, do you say use or utilize. Your favorite transitions. How you structure an argument, do you lead with the claim, lead with a story, lead with a question. Your punctuation habits, do you use em dashes, semicolons, sentence fragments, contractions. Your point of view, do you say we, or I, or you, and how directly do you address the reader. And critically, what your writing never does. The negative space matters.

Here's a pro move for the extraction prompt. Ask the model to ground every claim in a quote from your samples. For each trait you identify, cite the actual sentence that shows it. This forces the profile to be evidence-based instead of invented. Otherwise the model will happily make up flattering traits you don't actually have. Make it show its work.

And then, the step nobody wants to do but everybody needs to: you edit the extracted profile by hand. The model over-generalizes. It misses the soul of your writing, the weird specific thing that makes it yours. So you go in. You cut the generic stuff. You sharpen the vague stuff. You add the rules it missed. And you delete the flattering-but-false ones, the traits it invented to be nice. The model gives you a draft. You make it true. This edited artifact is the thing you reuse forever. And remember that foreshadow from the top? This is the artifact it was pointing at.

So let's lay out the anatomy of a complete reusable voice profile, so you know what you're aiming to end up with. Think of this as a fill-in template.

Start with a one-line identity. Who is this brand in a single sentence. Something like, a no-nonsense payroll tool for restaurant owners who hate paperwork. One line. The whole personality compressed.

Then three to five voice attributes, and here's the trick that makes them bite: pair each one with a we-are-X-not-Y boundary. Confident, not cocky. Plain, not dumbed-down. Warm, not chummy. That not-Y half is doing the heavy lifting. It tells the model where the edge is. Confident alone is vague. Confident not cocky draws a line.

Next, your Nielsen Norman coordinates. Where you sit on those four dimensions, as numbers or as words. Funny six out of ten, and so on.

Then structural shape. Your preferred opening move. Your sentence rhythm. Your paragraph length. How you like to end a piece.

Then vocabulary. Words and phrases you use. And words and phrases you never use. Both lists.

Then a blacklist. Five to ten patterns to avoid. This is your anti-slop list, and we're going to spend real time on what goes in it shortly.

Then two or three gold-standard examples of real on-brand writing. The few-shot samples live right here in the profile.

And finally, tone modulations. How the constant voice flexes for the contexts you actually face. How it sounds for a launch announcement versus an apology versus a how-to.

Now, if that feels like a lot, here's the starter minimum. The low-effort version that still works. Three to five bullets on how it should feel. One line on structural shape. A five-to-ten-item blacklist. And two or three examples you like. That's it. You can build that in twenty minutes and it'll already beat professional but friendly by a mile. Start there, grow it over time.

Okay, you've built the asset. Where do you put it so you're not re-pasting it into every single chat? This is where the current tools come in, and I'll walk all three, because they've actually converged on the same idea, which is the real lesson.

Quick gloss first. A system prompt, or custom instruction, is standing guidance the model reads before every reply. You set it once, and it applies to your chats without you re-pasting. That's the container we're looking for in each tool.

ChatGPT, from OpenAI. A few layers here. Custom Instructions are account-level standing guidance. There's a what to know about you field and a how to respond field, and it applies across all your chats. It's in settings, under personalization. Then there are Projects, which are workspaces with their own custom instructions that apply to every chat inside that project. This is the move for agencies and freelancers: one project per client or per brand. Each one carries its own voice. There's also Memory, which has approved saved memories plus implicit insights the model picks up about you. Memory is fine for personal preferences, but for consistent brand voice, use Custom Instructions instead, because Memory is fuzzy and it's not auditable. You can't easily see and control exactly what it's doing. And then there are Custom GPTs, which bake instructions plus reference files into a named assistant you can reuse and share. The way to think about the layering, broad to narrow: account Custom Instructions sit on top, then Project instructions, then a Custom GPT, then whatever you type in a one-off chat. Narrower wins for that context.

Claude, from Anthropic. Projects here are self-contained workspaces with persistent custom instructions, plus, and this is the big one, a knowledge base. You upload your writing samples and your style guide once, and every chat in that project can see them. So your few-shot examples just live there permanently. Claude also has custom Styles, which are reusable voice presets you can switch between mid-conversation. And there are two ways to make a Style. One, you upload writing samples and Claude analyzes and emulates them. Two, you describe the style in instructions. Notice what that first option is. It's the reverse-engineering technique we just learned, built right into the product. You hand it samples, it learns your voice. One caveat: Anthropic has been folding Styles into a broader personalization and Skills rollout, so the menu name might shift on you. Both approaches work as of now, but don't over-anchor on the exact label. Think of it as Claude's custom styles, whatever they're calling the menu this month.

Gemini, from Google. The container here is called Gems. A Gem is a saved, customized version of Gemini. You give it a persona, standing instructions, and up to ten reference files, which Google calls Knowledge. Build it once, open it whenever. Here's what's distinctive about Gems. They can connect to Google Drive. So instead of static uploaded files, you point the Gem at a live Doc, and when you update that Doc, the Gem sees the change immediately. Your voice guide becomes a living document that updates everywhere at once. And Google added sharing: from the Gem manager, you can share a Gem with view, edit, or copy permissions, which means you can spread one approved brand voice across your whole team.

So step back and look at all three together, because here's the durable, tool-agnostic takeaway. They all converge on the exact same shape. A saved container, which equals persistent instructions plus reference files or samples. ChatGPT Projects, Claude Projects, Gemini Gems. Same idea wearing three different hats. Once you own the voice profile, dropping it into any of these takes minutes. Which is the whole point. The asset is the lesson, not the tool.

If you want the quick comparison: Claude often gets cited as the most layered for voice work, because of Projects plus those sample-trained Styles. ChatGPT is the most ubiquitous, the default that everyone already has open. And Gemini is the most live, because of that Drive sync. Pick based on where you already work. The profile travels.

Alright. Now the pitfall. The one thing you will absolutely run into. The generic-AI tell. AI slop. This is the bland, telltale voice that screams a robot wrote this, and learning to spot it and strip it is half the battle.

First, why does it even happen? Because the model picks these words and structures since they're statistically high-probability, not because they're good. It's a statistical tic. The AI equivalent of a nervous habit. The model reaches for delve the way a nervous speaker reaches for um. It's not a decision. It's a default.

Let me give you the vocabulary blacklist. The tells. Listen for these. Delve. Leverage. Unlock. Harness. Unleash. Streamline. Optimize. Empower. Seamless. Innovative. Transformative. Landscape. Utilize. Cutting-edge. Paradigm. And there's a second wave: tapestry, testament, vibrant, intricate, showcase, bolster, garner, foster, underscore, navigate, holistic, multifaceted, pivotal, crucial, robust, nestled, boasts, remarkable, elevate, realm, ever-evolving. If you see a cluster of those, you're looking at slop.

Then there's verb inflation. This is when the model swaps a plain has or is for something puffed up. Boasts. Features. Serves as. Stands as. Represents. Marks. So instead of it's a hub, you get it serves as a major hub. Why? No reason. It just sounds more important to the statistical average. Plain copula verbs, is and has, are your friends. Use them.

Now the structural and phrasing tells, and these are sneakier than the words. The big one: it's not just X, it's Y. And its cousin, not only X but also Y. This is false-contrast, negative parallelism, and the models lean on it constantly. Then there's the throat-clearing opener: in today's fast-paced world, in today's rapidly evolving landscape. Just start. There's the avoidance of the simple verb we just covered. There's the rule of three everywhere, three adjectives stacked up to fake comprehensiveness: clear, concise, and compelling. There are the trailing significance clauses, the dash-i-n-g endings: highlighting its importance, underscoring the broader trend. Those tacked-on clauses that pretend to add meaning. There's hollow significance and legacy inflation: calling some mundane thing a pivotal moment, saying it marks a turning point, or leaves an indelible mark, when it's a software update. There's vague attribution: experts argue, studies show, with no named source anywhere. There are formulaic conclusions: despite its strengths, X faces challenges. And there are the sycophantic openers: great question, certainly.

Then the punctuation and formatting tells. The big litmus test here is em-dash overuse. A real human writer uses maybe two or three em dashes in a whole piece. The AI will drop twenty-plus. So if a paragraph is studded with them, that's a flag. Other tells: boldface scattered on every key term, like the whole thing is a textbook. Uniform paragraph and sentence length, no rhythm, everything the same size. And curly quotes or markdown artifacts leaking into pasted text, the little formatting crumbs that give it away.

Now here's the fair caveat, because I don't want to make you paranoid. Em dashes are a legitimate, lovely human device. So is the rule of three. The tell is overuse and uniformity, not any single instance. One em dash is fine. One trio of adjectives is fine. Twenty em dashes and a trio in every sentence is slop. Don't let slop-paranoia flatten genuinely good writing. The goal is your voice, not a sterile lab environment.

So how do you detect and strip it? Here are copyable scans you can run in two minutes. The em-dash scan: find every em dash, cut or convert most of them to commas or periods. The fifteen-worst-words scan: search each blacklist word and swap it for plain language. The not-just-X scan: search for not just and not only and kill those constructions. And the read-aloud test, which is the best one. Read it out loud. If it sounds like a brochure, or like a LinkedIn thought-leader clearing his throat, it's slop. Your ear knows.

But here's the upgrade that matters most. Don't just scrub after the fact. Bake the blacklist into your voice profile as an explicit never-do-this section. That way you prevent slop at generation time instead of cleaning it up afterward. The model reads the blacklist before it writes, and a lot of the junk never appears. Prevention beats cleanup.

Two nuances on slop before we move on. One, slop is a moving target. As the obvious tells get mocked publicly, em dash, delve, the model makers adjust their defaults, and your blacklist goes stale. So treat your blacklist as a living list. Refresh it from your own recent outputs, whatever junk keeps showing up. Two, don't over-correct into stilted. If you aggressively ban every flagged word, you can end up with awkward, hobbled prose that's contorting itself to avoid normal words. The goal is your voice, not maximum slop-avoidance as an end in itself. Stay loose.

Let me say a quick word about temperature, because it explains a lot about why chats feel the way they do. Temperature is a dial for randomness, for creativity. Gloss it like this. Low temperature, the model plays it safe, picks the most likely next word, comes out predictable and consistent. High temperature, the model gets more adventurous and varied, more likely to wander, and more likely to lose the thread.

The range runs from around zero to around two, with one point zero being the common default. Below one, you get focused and deterministic. Above one, you get creative but riskier, and past about one it can start to degrade, get incoherent. For brand-voice consistency, you want to lean lower. Steadier language, steadier format. Higher temperature is for ideation, brainstorming, when you actually want variety and surprise.

Now the honest caveat. Most marketers working in the consumer chat apps, the normal ChatGPT, Claude, and Gemini interfaces, don't get a temperature slider at all. It's exposed in the developer tools, the API, the playground, and some third-party apps. So two takeaways. One, understand the concept, so you know why a chat feels samey or, the opposite, why it feels like it's wandering off. Two, you can steer it indirectly even without the slider. Ask for five distinct options and you raise the effective variety. Say rewrite this staying very close to the sample and you push toward consistency. So you have the lever even when you can't see the dial. Just don't go around telling people they can type a temperature number into the chat box, because mostly they can't.

Okay, let me give you the complete workflow, start to finish. The spine. Seven steps. Write these down.

Step one. Gather samples. Three to five pieces, ideally building up toward three to five thousand words, of your best, most on-brand writing, across the formats you care about.

Step two. Reverse-engineer. Paste them in and ask for a reusable voice profile. Identity, three to five attributes with we-are-X-not-Y pairs, sentence and structure habits, vocabulary used and avoided, placement on the four Nielsen Norman dimensions, and a cited sentence for each trait.

Step three. Edit the profile by hand. Cut the over-generalizations, add what it missed, append the slop blacklist.

Step four. Save it persistently. Drop the profile into your tool's container, ChatGPT Custom Instructions or a Project, a Claude Project or Style, a Gemini Gem, plus the same three to five samples as few-shot examples. Or, better, generic-to-on-brand pairs.

Step five. Generate. Ask for the deliverable. Keep temperature low if you've got access to it, otherwise constrain it in the instruction: stay close to the samples.

Step six. Spot-check against the profile. Run the em-dash scan, the fifteen-words scan, the not-just-X scan. Check the output against your four Nielsen Norman coordinates. And read it aloud.

Step seven. Iterate, and this is the compounding one. When the output drifts, don't just fix that one piece and move on. Add the failure as a new rule or a new example in the profile. That way the fix sticks. It compounds. Every correction makes the asset permanently smarter, so you're not fixing the same drift over and over.

Let me make all of this real with a complete worked example, so you can see what a finished profile actually looks like. I'll invent a brand: Sprout and Co. Small-batch, indie software for freelance bookkeepers.

Identity, one line: practical, slightly irreverent software for solo bookkeepers who'd rather do the books than read a manual.

Attributes, with the boundary pairs: Plain, not dumbed-down. Dryly funny, not jokey. Confident, not salesy. On the reader's side, never above them.

Nielsen Norman coordinates: Funny six out of ten. Casual eight out of ten. Irreverent six out of ten. Enthusiastic five out of ten.

Structure: Lead with the payoff, not the windup. Short opener. Vary sentence length deliberately, a long one and then a short one to land it.

Vocabulary we use: you, contractions, concrete nouns, the actual feature name. Vocabulary we avoid: leverage, seamless, empower, solutions, robust, in today's, and exclamation points, max one per piece.

Blacklist: em-dash storms, not-just-X-it's-Y, rule-of-three padding, delve, unlock, trailing significance clauses, and great question.

Tone flex: Launch, playful. Billing or error, plain and a little apologetic, humor off. How-to, patient, second person, step by step.

Now watch what that profile does to actual copy. Here's a generic AI draft, the kind you'd get from a cold prompt. Quote: In today's fast-paced financial landscape, Sprout and Co. empowers bookkeeping professionals to seamlessly streamline their workflows and unlock new levels of productivity. It's not just software, it's a transformative solution designed to elevate your practice. End quote. Count the tells. In today's fast-paced. Landscape. Empowers. Seamlessly streamline. Unlock. It's not just X, it's Y. Transformative solution. Elevate. It's a slop bingo card.

Now the on-brand rewrite, run through that profile. Quote: You didn't become a bookkeeper to fight your software. Sprout does the boring parts, reconciliations, reminders, the chase for missing receipts, so you can close the month and get your weekend back. No forty-tab dashboard. Just the numbers, done. End quote.

Hear the difference? What changed. We killed in today's fast-paced, empowers, seamlessly streamline, unlock, transformative solution, elevate, and the not-just-X-it's-Y construction. And we added second person, you. Concrete nouns, reconciliations, receipts, the forty-tab dashboard. A short landing sentence to close. And the brand's dry register throughout. Same product. Same basic message. One sounds like every company's AI. The other sounds like a company you'd actually want to buy from. That's voice transfer. That's the whole game.

Last piece. Why this matters beyond any single email. Two kinds of consistency.

First, consistency across formats. Your voice should survive a format swap. A tweet, an email, a help-doc, and a landing page should all feel like the same brand. Which is exactly why, back in step one, you gather samples across formats. If all your samples are blog posts, you'll accidentally build a profile that's blog voice only, and your help-docs will sound like a stranger.

Second, consistency across people. Your voice should survive a person swap. When the voice is written down in a shared profile, a new hire, a freelancer, and the AI can all produce on-brand copy without the founder hovering over every draft. That's the real payoff of documenting it. The voice stops living in one person's head, where it's a bottleneck, and becomes something repeatable. This is precisely why a written, shared profile beats voice-in-someone's-head.

And the tools support this now. Gemini Gems are shareable with view, edit, and copy permissions. Claude and ChatGPT Projects are shareable within a workspace or team plan. So one approved profile propagates across the team instead of everyone quietly improvising their own version. And with Gemini's Drive sync, you get a truly living document. One canonical voice guide, in a Doc, updating everywhere at once the moment you edit it.

The model to imitate for all of this is the one we started with. Mailchimp published their style guide for the world to read. Voice, tone, grammar, and mechanics, all documented in one place, written so anyone can pick it up and apply it. That's the standard. Not because it's fancy, but because it's usable.

So, where that leaves you. The default AI voice is bland corporate optimism, and it's strong, so you have to actively override it. You override it with repetition: a written spec, real examples, and low temperature where you can get it, all pulling in the same direction. You override it with examples over adjectives, because examples carry the information adjectives lose. And you override it by building the asset once and grounding the model in it every time, because the asset is the value and the prompt is just the delivery. Build the profile. Edit it by hand. Save it where it persists. And feed it back in, every single time.