Writing the Brand-Voice Profile

The founding charter

The core problem: brand voice instructions usually collapse into adjectives — friendly, bold, professional — which an assistant can't actually act on, so it hands back the generic average of everything anyone has ever called "friendly." The fix is to build a one-page voice profile made of behaviors instead: rules specific enough that a competent stranger could follow them without asking a question.

Building it starts with gathering three to five pieces of writing you already know are good — published, approved, human-written, ideally the same format. New brands with no back catalogue can dictate five minutes of talk and transcribe it instead; borrowed agency copy, template text, or old AI drafts don't count as evidence of voice at all.

From there, an assistant is used only to read the samples and report observable patterns — sentence length, where paragraphs land their point, repeated words, things the writing avoids — with the rule that nothing gets stated unless it shows up in at least two samples. That draft output then gets edited in four passes: cut compliments, rewrite each surviving line as a checkable behavior, rank the load-carrying rules to the top, and add a short never-list of banned words and tics. Two or three whole samples get pasted at the bottom, since a shown pattern beats a described one.

The finished page slots straight into the voice field of a six-part prompt template, then gets tested by holding back one real piece, running the same brief, and comparing openings and closings to find which specific rule the draft broke — each miss becomes one more line.

The chapter's pitfall: profiles that keep growing past the point of usefulness. Once a document balloons into thousands of words of brand philosophy, rules buried in the middle get skimmed and drafts quietly go generic again. The fix is subtraction — keep only lines that a competent writer wouldn't produce by default — while a longer version can still exist separately for onboarding humans.

Consequential AI-marketing news and practical shortcuts

Anthropic has turned on memory across regular Claude chats and Cowork, with remembered items grouped into inspectable, editable topics; sensitive categories like health are excluded by default. Memory arrives on by default for free, Pro and Max plans, but off by default for Team and Enterprise accounts. Anyone running a voice profile is pointed to check that memory list and delete anything about writing style that contradicts the profile, before testing it.

Separately, Google has introduced Studio Skills inside its Workspace side panels — saved, shareable bundles of rules, templates and reference documents for on-brand drafting — and widened hands-free drafting and Drive/Gmail-grounded lookup to its paid AI tiers. The suggested move: share the finished voice profile with one colleague doing customer-facing writing to see whether the rules hold up for someone other than their author.

You have a prompt template saved now. Six fields, sitting in a note or a doc, ready to be pasted and edited for whatever the day throws at you. And you have one correction banked into it — a line you added after an assistant got something wrong, so it never gets that thing wrong again in quite the same way.

So here is a question worth asking before you write anything else this week. Which of those six fields do you retype the most?

For almost everyone, it is voice. Audience changes per task, and it should. Goal changes per task. Format changes. Constraints mostly change. But voice does not change. Your brand sounds the way it sounds whether you are writing an email to a practice manager or a caption for a product photo. And yet every single time you open a fresh chat, you sit there and describe it again from scratch. A few sentences about how you write. Maybe a sample pasted in, if you can find one fast. Then you do it again tomorrow, slightly differently, because you were in a hurry and you left out the bit about not opening with a question.

That is the repeated work this chapter ends. By the time we are done, you will have one document — about a page — that you paste into the voice slot and never rewrite. This chapter is about how to build it out of things you already own, and how to prove it works.

Voice is a set of behaviors, not a bundle of adjectives

Start with the reason most voice instructions fail.

Ask a marketer to describe their brand voice and you get adjectives. Friendly. Professional. Confident. Approachable. Bold but never arrogant. Human. Those words are not lies. They are just not instructions. Think about what an assistant can actually do with "friendly." It has read an enormous amount of text that somebody, somewhere, would call friendly. So it produces the average of all of it. That average is exactly the generic, smoothed-over copy you are trying to get away from. You asked for friendly and you got everybody's friendly.

Now compare that with a different kind of sentence. We name the thing before we explain why it matters. We never open with a question. Our sentences run short, and when one runs long it is because it is carrying a list. We say "you" and almost never "clients." We do not use the word "solution."

Every one of those can be copied. You could hand them to a new freelancer on their first morning and they could follow them without asking a single follow-up question. That is the test for a line in your voice profile: could a competent stranger follow it without needing to interpret it? If yes, it belongs. If it needs interpreting, it is an adjective in a longer coat.

There is a second reason to work this way, and it matters more than it looks. Adjectives are self-description, and self-description is unreliable. Most brands think they sound punchier than they do. Most founders think they sound warmer than they do. If you build a voice profile out of what you believe about your writing, you will build a profile of a brand that does not exist, and then you will spend weeks wondering why the drafts feel off. So we are not going to ask you what your voice is. We are going to look at what your writing actually does, and write that down.

Which means, before the workflow starts, you have to go and get something.

You need three to five pieces of your own writing that you already know are good. Not good in theory. Good in the sense that they went out, somebody approved them, and you would be happy to see them again. Published work, or work that cleared a review. Written by a person, not generated. And ideally the same type of asset, or at least close, because a landing page and a launch email do genuinely behave differently and mixing five wildly different formats gives you a blurrier picture than picking three that rhyme.

Three is enough. Five is comfortable. Ten is worse than five, and we will get to why later.

Now the two situations where you cannot do that, because plenty of listeners are in one of them.

The first is that you are new. Brand-new brand, or a freelancer taking on a client with no back catalogue, or a founder who has genuinely never written the marketing copy down because it has only ever lived in their head and in sales calls. You cannot extract a pattern from nothing. So make something to extract from. Write one short piece by hand — one email, one page of a pitch, four hundred words is plenty — and write it badly if that is how it comes out, as long as it comes out as you. Or better, and faster: get the founder talking for five minutes about what the product does and who it is for, record it, and transcribe it. Any assistant will turn an audio file or a rough transcript into readable text. What you now have is unpolished, repetitive, and full of half-finished sentences, and it is still worth more than a beautifully written imaginary voice, because it is real. A rough real voice on the page beats a polished invented one every time, because the invented one has nothing under it and you will drift away from it within a month.

The second situation is the trap. You do have samples — a stack of them — but they came from somewhere else. An agency wrote them. They came out of a template library. Or, most common now, they were drafted by an assistant a few months ago and lightly edited. None of those are evidence of your voice. Agency copy is the agency's voice with your product names in it. Template copy is nobody's voice by design. And assistant-drafted copy is the exact generic average you are building this profile to escape, so feeding it back in teaches the machine to imitate itself. That is worth being strict about, because it fails quietly. The profile will look fine. The drafts will keep coming out bland. And you will blame the tool.

So: three to five real pieces, written by humans on your side. Get them into one place before you open a chat.

Here is what you do with them.

Paste the samples in, and then ask the assistant to read rather than write. That distinction is the whole move. You are not asking for a style guide. You are asking it to report what it can see. Something close to this: here are four pieces of writing from the same brand, ask it to describe the patterns it can observe across all of them — typical sentence length, how paragraphs tend to open, whether the point arrives at the top of a paragraph or at the bottom, which words show up repeatedly, what the writing consistently avoids doing — and to only state patterns it can point to in at least two of the samples.

That last clause is doing real work. Without it you get flattery. With it you get evidence.

What comes back will be a mix. Some of it will be sharp and genuinely useful — the observation that your paragraphs almost never open with a subordinate clause, say, or that you use the customer's job title rather than the word customer, or that you never end on a call to action but always on a consequence. Some of it will be filler. And a good chunk of it will be compliments, because these tools are built to be agreeable. You will get lines like "the writing is warm and authoritative" and "the tone strikes a confident balance." Those are the adjectives coming back in through the side door.

So that output is a draft. It is not the profile. The profile is what you make of it, and the editing is your job, not the assistant's. Four passes, and they go quickly.

First, cut every line that is a compliment rather than an instruction. If the line makes you feel good and tells a stranger nothing, delete it. "Warm and authoritative" goes. Be ruthless here, because these lines are the ones that make a document look complete while doing nothing.

Second, rewrite each surviving line as a behavior. The assistant will often hand you an observation in a slightly useless shape: "sentences tend toward brevity." Turn that into the thing to do. Most sentences run under twenty words. Or, better, because it is more specific: no sentence carries more than one idea. If a line describes something your writing does not do, write the refusal down flat. We never open with a rhetorical question. We do not use the word journey.

Third, rank them, most important first. This one gets skipped and it costs people. When you paste a long block of instructions in, the assistant does not weigh them evenly. Rules sitting at the bottom of a long document compete with everything above them, and they tend to lose. So put your two or three load-carrying rules at the top, and let the small stylistic preferences sit underneath, where it does not matter much if they slip. If you only get three rules honoured, you want to have chosen which three.

Fourth, add a short never-list. Words your brand does not use. Phrases that make you wince. Structural tics you keep having to edit out — the three-word sentence fragment for emphasis, the paragraph that ends by restating its own first line, the sign-off that thanks the reader for their time. Refusals are unusually easy for a machine to follow, because there is no judgement involved. It either used the word or it did not. A never-list of eight items will change your drafts more than a page of aspiration.

Then, at the bottom of the document, paste two or three of the original samples in whole. Not extracts. Whole pieces.

This is the same lever as the examples field in the prompt template, and it is worth saying plainly why it earns space here. A pattern shown beats a pattern described, and it beats it by a wide margin, because a sample carries a hundred small decisions you never managed to write down as rules. Rhythm. Where the paragraph breaks land. How much you explain before you assert. You could not itemise those if you tried. You do not have to. You just show them. The difference between this and what you were doing yesterday is that yesterday you went hunting for a sample every time, and today it lives in the file.

Cap the whole thing at roughly one page. That includes the samples. If the samples are long, use two rather than three, or trim each to its first few hundred words, since the opening of a piece carries more voice per line than the middle does.

That is the document. Now use it.

It goes in the voice slot of the six-part template. That is the entire integration step, and its plainness is the point. You paste the profile in where you used to type a few sentences, and then you fill out audience, goal, format, constraints and examples the way you already do. Nothing else about your workflow changes. What changes is that the voice field is now the same every time, which means when a draft comes back wrong, you know it was not because you described your voice differently on a Tuesday.

Worth flagging, so you know you are building something that keeps paying out: this file has a longer life than this chapter. Later in the series, when we load your brand into a saved assistant so it stops guessing about who you are, this profile is a large part of what gets loaded. And much later, when work gets handed to something that runs multiple steps on its own, this is the brand context it holds while it works. You are writing it once for the prompt template. You will end up using it three times over. So it is worth an hour now rather than twenty minutes.

Then prove it. This takes about ten minutes and I would do it today rather than trusting the document on sight.

Hold back one real piece you did not feed in. This is why you want a piece in reserve — if you have five good samples, use four to build the profile and keep the fifth sealed. Now brief the assistant on that same task, using your six-field template, with the profile in the voice slot. Same audience, same goal, same format, same constraints. Do not tell it what the real piece said. Then read the two side by side.

The instinct is to score it. Did the AI win. That is the wrong question and it will teach you nothing, because it did not win and it was never going to. The right question is narrower and much more useful: which specific rule did the draft break?

That question always has an answer. Maybe the profile said the point arrives first, and the draft spent a whole paragraph clearing its throat before it said anything. That tells you the rule was too abstract, and it needs to become something a stranger could check — the first sentence states what the reader gets, not why the topic matters. Maybe the draft used a word you hate that you had never noticed you hate, in which case it joins the never-list. Maybe the draft is technically obedient and still feels wrong, which usually means the rule you actually need is not in there at all, and comparing the two pieces closely enough will surface it. Read the two openings against each other, then the two closings. Those are where voice sits thickest.

Every miss becomes one more line. That is the loop, and it is the same fix-the-field habit you already started when you banked your first correction into the template — the profile is just a better place to bank the corrections that are about voice, because those are the ones that recur on every single task.

The loop has one dangerous property, though, and it is the pitfall for this chapter.

The style bible nobody follows

The failure is not that people never write a voice profile. It is that they write one and then keep feeding it.

Here is how it goes. The document works. So you add to it. A miss becomes a line, which is correct. But then a preference becomes a line. Then a mission statement gets pasted in, because someone in the room says the voice should reflect the values. Then somebody adds a section on the em dash. Then the tone matrix arrives, with different guidance for eight channels. Six months on you have four thousand words of brand philosophy, and it is honestly quite an impressive document. Your team is proud of it. It reads beautifully.

And your drafts are worse than they were at page one.

The mechanism is worth understanding, because it is not the machine being lazy. When you hand an assistant a long instruction document and a request, it does not treat all of it with equal weight. In practice, what sits closest to the actual request gets followed hardest, and the material further away gets thinner attention. Your rule about never opening with a question, sitting on paragraph nineteen of a wall of aspiration, is competing against three thousand words of other material and losing. Nothing announces this. There is no warning that half your document is being skimmed. The drafts just quietly come out generic, obedient to the two or three things that happened to sit near the top or near the end.

The recognition signal is simple, and once you know it you will spot it in other people's setups too. The drafts stop improving even as the document grows. You add a rule, the next draft is no better. You add three more, still no better. If effort is going in and quality is flat, you are past the point where the document helps, and every line you add from here makes it slightly worse.

The fix is subtraction, and it is uncomfortable because you have to delete work.

Go through the profile and ask one question of each line: does this distinguish us from any competent writer? "Write clearly." A competent writer already does. Cut. "Avoid jargon." Everybody says that, and it is not a rule, it is a hope. Cut. "Be helpful, not salesy." Cut, unless you can turn it into something checkable. But "we name the mechanism before the benefit" — a competent writer might do that or might not, so it stays. "We never use the word partner as a verb" — nobody would guess that, so it stays. What survives is usually somewhere between five and a dozen lines, and it will feel thin, and it will outperform the four thousand words. Let the pasted samples carry everything you deleted, because they do carry it, and they carry it without spending your document's attention budget on prose about prose.

Then, so the deletion does not feel like loss: keep the long version. Nobody says you have to throw it away. It is a fine document for onboarding a human, who will read it once and absorb it, and who can hold four thousand words of context in a way this technology does not. It is simply not the thing you paste. Two files, two jobs.

One more thing, since we have been using the clinic all the way through — that short email to practice managers about the text-reminder feature. When you run the ten-minute test, keep the task's own conditions attached to the task. That email had to work on a phone screen, which is why its whole argument has to land in the first line. Note that in the format field, where it belongs, because it is true of that email and not of your brand. It is not a voice rule. If you push it into the profile as one, you will get punchy opening lines on your long-form pages for the next year and slowly wonder why everything reads like a notification. Constraints that come from where a piece is read belong to the brief. The profile holds only what is true of your writing everywhere.

So: three to five real samples, an assistant used as a reader rather than a writer, your edit turning observations into rankable behaviors, a never-list, two whole samples at the bottom, one page, and one held-back piece to test it against. That is an hour of work that ends a small piece of retyping forever and quietly builds the thing your brand will run on later.

Two things landed in the last couple of weeks that sit close enough to this to be worth your attention, both about where brand rules get stored rather than how they get written.

At the end of August, Anthropic turned on memory across regular Claude chats and its Cowork product, and — the part that matters for this chapter — made it inspectable. Under the memory settings you can see the remembered items grouped into topics, and edit or delete them one at a time: preferences, instructions, traits it has picked up about how you like things done. Sensitive categories like health and personal beliefs are held out of memory by default unless you switch that on. The default state differs by plan, though. On the free, Pro and Max tiers memory arrives switched on. For Team and Enterprise organisations it arrives switched off, which means if you are on a company account and expecting it, check before you assume.

Why a marketer cares: this is a fourth place your voice instructions can now live, alongside the profile you paste, a saved workspace's standing instructions, and reference files. That sounds like a convenience and it is also a way to lose track of what is steering a draft. If a draft comes out wrong and you cannot tell whether it followed your pasted profile or something it remembered from a chat three weeks ago, you have lost the thread you spent this chapter learning to hold. Smallest next action: open the memory topics list, read what is in there, and delete anything about your writing style that contradicts the profile you just wrote. Five minutes. Do it before you test the profile, not after, or your ten-minute test is measuring two things at once.

The second item, from a day later: Google announced something it calls Studio Skills inside the Workspace side panels — the assistant panel that sits beside Docs and Gmail. A skill is a saved bundle of rules, templates and reference documents that you can create once and share with colleagues, aimed squarely at routine drafting that has to stay on-brand. A few days after that, Google widened its Docs Live and Gmail Live features — hands-free drafting and conversational lookup grounded in your Drive and Gmail — to its paid AI tiers.

Why a marketer cares: a shareable bundle of rules plus reference files is, structurally, the profile you just built, with distribution attached. If your team drafts inside Google Docs and Gmail all day, that is an argument for the profile living where the writing happens rather than in a personal note only you can find. Smallest next action, and this is genuinely small: after you have written and tested the profile, share the file with one colleague who writes customer-facing copy and ask them to use it on their next task. You will find out fast whether your rules are followable by someone who is not you, which is the test that matters and the one you cannot run alone.

Both of these are this week's picture rather than a verdict. Storage locations for brand rules will keep multiplying, the defaults will keep moving, and the document you wrote today works in all of them.