Brand Voice Is Language + Cognition

Most brand-voice guides capture only the words, which is why AI writing built on them still sounds generic. This is how I rebuilt ours from the transcripts of three people on our team, and what I found underneath the language.

For the umpteenth time, my producer, our team’s heaviest hitter for writing, came to me with a voice problem. My plate was full and I’d been ignoring it.

“Elizabeth, this blog post is not great.”

He was right. I’d stripped out the em dashes and words like “quietly,” and it still sounded like AI, and worse, it didn’t sound like us. I told him I’d look at it later, already planning to push it to next week. Then I opened an email we’d sent that morning, and the subject line was a cardboard cutout of our brand, flimsy and two-dimensional. That was the moment I stopped pushing it to next week.

I knew what I’d been doing, and it wasn’t working. I’d been telling our AI what not to do: no em dashes, no rhetorical questions to open, not this word, not that phrase, not the tic that kept showing up. Every time I closed one gap it opened another. I told it to stop hedging and it started over-claiming. I told it to stop over-claiming and it hedged. Maddening.

The list of what not to say

A real person doesn’t talk from a list of permitted and forbidden words. They talk the way they think, and the words are the part you can hear.

My style sheet only had the part you can hear. So I stopped writing rules and went to the source.

We had hours of webinar footage of our founding team, which meant I didn’t need Claude to guess what our voice was. I had it on tape. I pulled the transcripts of three people in particular.

One of them sees the whole board and can sell a vision. One knows the research all the way down and won’t let you approximate it. One is a coach, warm and practical, always pulling it back to what a teacher does on a Monday.

The mission underneath was the same, but each of them had a distinct way of thinking and speaking. I wanted the blend. One person’s voice is just a copy, and my hunch was that braiding three complementary ways of thinking would produce something none of them would have written alone.

Three people, one habit

I had Claude run the analysis and two things came back that I hadn’t seen coming.

The first was how much depth there was. Each person didn’t have one way of answering a question; they each had several, and which one they reached for depended on the kind of question: one answer for a factual gap, another for a disagreement about values, another for a half-formed idea that needed room.

A voice guide takes a person and flattens them into three adjectives and a list of approved words.

The second thing ran deeper. All three of them spoke differently, but they had a few habits in common, and the clearest was how they handled the limits of what they knew.

Generic AI hedges as a reflex, softening sentences while over-claiming in the same breath. None of the three did that. They were precise about the limits of what they knew, and they would each open the same honest way, marking what they knew against what they believed: “we know comprehension matters for reading, and we believe social studies gives students the knowledge that unlocks it.”

Then each of them turned that limit somewhere different. One asked what question we’d need to answer to know if it was true. One treated it as something we’d get to find out. One pointed past it to a bigger possibility.

Same claim, same honesty about where it stopped, and each of them turned the stopping point into an invitation.

How they thought

At that point I stopped treating it as a language problem. How a person handles the edge of what they know happens before the sentence does. The words are where it surfaces.

So I asked Claude to go down another level, past the language to what was distinct about how each of these people thinks.

This was the layer I hadn’t bet on. All three spoke from experience in the work, not in generalities. They all said the hard part before the solution: “here’s where implementation fails” before “here’s how you teach this curriculum,” and “I’m not going to sugarcoat it, doing it this way is harder” before “we can get every kid writing with purpose.” Same habit, same place in the sentence, and none of them had ever talked to each other about how they talk.

I could ask for stronger verbs all day. I could tell the model to stop writing in threes. None of it made our writing better or truer to our brand, because it was missing half the puzzle.

Feeding it words and rules didn’t work, and it shouldn’t have. Nobody learns a language from a list. We learn from the people around us, and we pick up how they think while we’re picking up how they talk, and both keep changing as our lives and our relationships change. Your voice is never only yours, and it’s never finished. It’s a blend you’ve been assembling without noticing since before you could read.

Same words, different order

I had Claude write the same blog post two ways, once with our language rules alone and once with the language and the cognition patterns together.

The language-only version opened like this:

A first grader stares at the blank page. She learned something today. She talked about it, listened to her classmates, built an idea. But the page stays blank. This is easy to read as a writing problem. It usually isn’t.

That already reads like a person wrote it. Someone is in the room, the sentences are plain, the abstraction is gone. A year ago I would have called it done. But it hands you the right answer, that it usually isn’t a writing problem, and keeps moving. It never sits with you in the wrong answer first.

The version with cognition added opened almost identically:

A first grader stares at the blank page. Her teacher asked her to write about what she learned, and she has learned plenty. She listened, she talked, she built something in her mind, but the page stays blank. This is one of the most common scenes in early elementary classrooms, and it is easy to misread. The blank page looks like a writing problem. It is actually a content and purpose problem.

Read the two openings against each other. The words barely changed. What changed is the order, the timing, and one shift of perspective.

The cognition version says the misread out loud before it corrects it. “It is easy to misread” tells the reader who has spent three years explaining the blank page as a motivation problem that she isn’t about to get called out. She’s about to get seen. That’s one of our three voices, the one who can diagnose a situation without ever diagnosing the person.

Then it shows you the wrong perception on purpose before swapping it. “The blank page looks like a writing problem” sits there for a beat before “it is actually a content and purpose problem” replaces it. The language-only version just gave you the right answer. The cognition version lets you hold the wrong perception first, because it’s hard to give up an idea you haven’t looked at yet. That’s the second voice, the one who always shows you the idea already in your head before offering a better one.

And in the body, the cognition version calls the problem what it is, which turns a description into something you can say in a meeting. That’s the third voice, the one who can’t help moving from principle to example to what it means. The language-only version describes the same problem without ever calling it anything, so it stays abstract.

You can’t get any of that from a word list. The difference is whether the writing acknowledges where your head is before it tries to move you, and that comes from three people’s way of thinking, braided into one.

The people whose work it is

A while later I ran the same process on myself. I fed in three months of my meeting transcripts, every Slack message I’d sent, and my writing on creative briefs, and asked the same questions I’d asked about the three of them. What came back described someone who narrates into a problem instead of leading with the answer, someone who uses people’s names mid-thought to think out loud. It caught the thing I do where I end a sentence on “so” and just trust you to finish it. (My husband has been pointing this out for years. He did not get a citation.) I’d never once described my own voice that way, and it was more accurate than anything I could have written about myself.

Where it goes in the process now

Something this good at capturing how a person thinks can counterfeit one just as easily, so we decided where it’s allowed to work and where it isn’t.

Before anyone drafts, it reads. Transcripts, old posts, the founders’ webinars. It finds patterns across forty hours of people talking better than I can, and it doesn’t get bored on hour thirty.

Then it writes the first draft, with the voice system running. That draft is a starting point. It is not what we publish, and everyone on the team knows it. Scoping a campaign brief used to take me an hour. Now it takes ten minutes of reading.

The edit is where the person comes in, and it’s the person whose work it is. My producer reads for whether it sounds like us, and the cognition layer is where it still gets caught most often. It hands you the right answer without sitting in the wrong one first, and putting that beat back is something only a person who knows the reader can do.

Some things it never writes alone. A subject line. A headline. The sentence that says what we believe. Those carry a promise, and a promise goes out with a person’s name on it.

The last check is the oldest one. Read it out loud. If you wouldn’t say it in a meeting, it isn’t ours yet.

Most people I talk to agree on the basics. You need strong thinking before you write anything, AI can be part of your writing process, and it still shouldn’t produce the final piece. We ended up in the same place. In our process the draft is a starting point, and a person decides what goes out.

I’d add one more thing: the system is never done. A voice comes from how people think, and people keep changing as their work and their relationships change. The three people I analyzed will think differently a year from now, so the analysis has a shelf life, and I plan to run it again.

If you want to try this on your own team, start with what you already have. Pull the transcripts from a few webinars, calls, or long Slack threads, and read them for how each person handles the edge of what they know. Notice what they do right after they say “we don’t know yet.” You’ll hear something no style guide would have given you.