40 years of storytelling craft, now built into the systems that keep a voice consistent.
Media spend bought the attention, and the writing only had to not get in the way.
That has changed. Ad costs have risen structurally, targeting has degraded, and people scroll past. So brands are moving to the channels they own — email, WhatsApp, organic, community — where attention cannot be bought at all.
Nobody is obliged to open your email. On an owned channel the only thing that works is being worth reading.
The obvious answer is to write better copy. But better is not a decision you make once.
Every email, every page, every caption is a fresh judgment, made by whoever is writing that day. A founder can hold that line at three pieces a week. At thirty they cannot.
Write the voice down and it becomes one decision, made once, applied every time — by a new hire in their first week, or by the AI drafting at volume. You stop being the review layer.
What separates you is not deciding to sound different. It is being the only one who wrote down what different means.
The reason you cannot simply write your way out of this is that most of the volume is no longer written by hand. It is generated — and a generator produces the average of everything it has read unless somebody tells it otherwise.
Not a document that sits in a drawer. A working system your people run every day.
I define your voice as an operational specification — attributes, tone rules, the words you use and the ones you don't, and each trait shown in your own lines. Whoever writes for you, by hand or in your AI tools, starts from the same place, so the volume you're producing starts sounding like you. In a Pilot, you install it yourselves. In a Complete Build, I install it for you.
The old fix doesn't apply. A brand guidelines PDF was written for humans to read. An AI can't generate from a mood board — it needs the voice supplied as something it can work from. Without that, what comes out is competent, fast, forgettable.
The rare part isn't the AI. Knowing what makes a voice distinct — and codifying it so a machine can hold it — is the harder half. It's the half AI can't supply on its own. That's what forty years of storytelling craft is for.
Used well, AI doesn't flatten a voice. It multiplies it.
You built something with a point of view, then scaled the content and watched it go quiet. The point of view didn't disappear — it just never got written down in a form the tools could use.
Your voice is the product, and scaling it without losing it is the whole problem. AI can carry more of it than you'd think — once it's been told what it's carrying.
You find everyone else's voice for a living. Your own deserves the same rigour, and the tools your team already uses can hold it once it's properly specified.
You can start with the Pilot or the Complete Build. The Retainer comes after either one, if you want it.
One brand voice, defined and made AI-ready. Low risk, fixed price — a real look at the work before any bigger commitment.
You install the kit yourselves. The setup steps for each tool you use are in the kit.
Everything the Pilot delivers, extended across your whole content surface — every format, channel, and use case your brand speaks in — and installed for you.
A built system drifts — new people, new channels, new tools. The Retainer is continued help, by choice, once the build is done. After a Pilot, it's audits and judgment:
On pricing: the Pilot has a standard fixed fee, not priced per brand, and you get it in writing after our call. The Complete Build is quoted once I've seen your brand: a fixed fee for a stated scope, with anything added to that scope quoted separately. No hourly meter, no open-ended scope. You work directly with me. The Pilot is a deliberately accessible entry point. If you want continued help afterwards, the Retainer is there. It's your choice, not a condition.
Not every engagement includes all five.
Your voice written as rules and worked examples from your own material — attributes, tone by channel, the words you use and the ones you don't. Everything else runs on this.
Instructions to paste into your AI tool, with both Brand Voice System documents attached. The instructions are the same whoever is typing. The voice comes from the documents, not from how well someone asks.
The kit installed and connected inside your own tools and accounts, with any subscription or access held in your name. Nothing of mine keeps running afterwards. Complete Build only.
A working session with the people who'll run it, and a one-page run sheet.
Your published output read against your Brand Voice System: what's drifting, and why. It's advice, and your team makes the fixes. Part of the Retainer after a Pilot.
That decade taught me the thing all of it rests on: how a story makes someone feel something, and stay.
My track record is the craft. I've also tested the AI side on my own brand, twice, on two different days. Same request, two ChatGPT chats: one with my Brand Voice System, one without. Both times, the one with it kept to my brand voice. The one without didn't.
Latest issue: Isn’t that just a prompt?
One request, two chats, one difference: a written voice. See what it changed in an email your brand could send.
Name the brand. Before the call, I personally read your public pages. On the call, I ask a few questions, and you tell me what concerns you. The fee follows the call, in writing.
If you already have guidelines, say where to find them when you book. That's optional.
One request, two chats, one difference: a written voice. See what it changed in an email your brand could send.
Facts alone don't carry a voice. One request, run with and without a voice system, shows what changes for your brand.
Much AI copy sounds alike. Why your brand needs a voice system, shown each issue: one request, with and without one.
Much AI copy sounds alike. Why your brand needs a voice system, shown each issue: one request, with and without one.
I build voice systems so that a brand's writing sounds like the brand, and in this newsletter I show what that changes, one request at a time.
A voice system is a brand voice written down as a set of documents a brand writes from: how it sounds, the words it uses and the words it avoids, and a few of its own best lines to measure new writing against. A person can work from it, and so can a tool.
For this first issue I made up a brand: a family coffee farm in Kona, Hawaiʻi. The farm is fictional, and this is my own demonstration, not client work. I asked for one page, the farm's "Our story" page for its website, up to 400 words. I ran it in ChatGPT on 16 October 2026, in Think mode, in temporary chats with personalization off, so no saved memory or settings of mine shaped the answers. The voice system is plain documents, which any AI tool can be given. This is one run. What I say changed, changed here.
The first chat got the request with two files:
Neither file says anything about how the farm should sound.
The second chat got the same request and the same two files, plus two more. Those two are the voice system:
The difference shows straight away:
The page still had slips against the facts:
I read the page against the facts and sent one correction. It named what was wrong and pointed back to the files. It didn't supply a single word of the fix.
The revised page fixed every slip against the facts that I named, including those three:
What else changed: it opens on what the page is for, the work done on the farm since 1983. The deck "still has its job". The old building "no longer mills coffee".
Fixing those loosened other things:
It still reads as a timeline, without the rhythm the voice system asks for.
The real work is in the voice system itself. In this series I leave it untouched, so you can see what it does on its own, untested. For a brand, it is tested on the brand's own work, and what the testing finds is corrected in the voice system itself before it is handed over.
Now think of the last page your brand put out with AI's help. Could its closing line have come from any brand in your line of work? In my view, the page without the voice system ended on a line any farm could use, and the page with it began to sound like this farm. Does yours sound like you?
Series note. The voice system in this series is deliberately left uncorrected so readers can see the process. A client's voice system is tested on their own real tasks and corrected before handoff.
Facts alone don't carry a voice. One request, run with and without a voice system, shows what changes for your brand.
You've scrolled past captions that could belong to any brand. So have your customers.
In a Clutch survey of 601 consumers in August 2026, 86% said they had come across AI-generated content on social media. The question isn't whether AI was used. It's whether the caption sounds like you.
In issue 1, I showed what a voice system is: a brand's voice put in writing, with the words it uses and avoids, and a few of its best lines.
This time nothing changed except the job. Same fictional coffee farm, same facts, same law, same voice system, not a word altered. It's my own demonstration.
A voice system has to hold across everything a brand writes. So this time I asked for Instagram: five captions about the harvest, up to 40 words each, one of them inviting people on the farm tours.
I ran it on 20 October 2026, in ChatGPT, in Think mode, in temporary chats with personalization off. Nothing of mine could colour the answers.
It's one run. What I point out is what happened here. Each issue adds one more run to the record.
The first chat had the facts and the law: what's true about the farm, and what it may and may not claim, compiled from public sources, not legal advice. Nothing on how the farm should sound.
The second chat had the same request and the same two files. It also had the voice system: one file on how the farm sounds, and one with four of its best lines to measure against.
The gap is smaller than on the story page in issue 1. It still shows:
The captions also slipped on the voice system's own rules:
I sent the second chat one note. It listed the five slips and pointed back to the files. The rewriting I left to the chat.
What got better:
What didn't:
My verdict, two issues in
In my view, across both issues, each chat was consistent. That's the first thing I noticed.
Without the voice system, the writing was generic every time. A brochure with headings and a line about generations on the story page. Padding and an exclamation mark in the captions.
With it, the writing was the farm's every time. Plain and particular, on a web page and on Instagram. The voice files are what changed it.
That is the point of a voice system. A brand doesn't write one page. It writes captions, emails, signs, job ads. The voice has to be the same in all of them.
The gap was smaller this time, and that tells you something:
Two things the voice system hasn't done yet:
In this series I don't correct the voice system, so each issue shows it as it is. A client's voice system goes through those rounds on the client's own writing. Each slip a test finds is corrected in the files before the brand takes them over.
Scroll back through your brand's last five captions. Take your brand name off one. Could a follower still tell it was yours?
Series note. The voice system in this series is deliberately left uncorrected so readers can see the process. A client's voice system is tested on their own real tasks and corrected before handoff.
One request, two chats, one difference: a written voice. See what it changed in an email your brand could send.
You've probably tried it. You asked a chat for an email, read what came back, and thought: a better prompt would fix this.
So when I say a brand needs a voice system, the fair question is the one in the title. Isn't that just a prompt?
This issue answers it with something you can inspect. It's the same fictional coffee farm as the first two issues. This time the job is an email to its subscribers: thank them for staying, give the roast and ship days, say how to change, pause or stop. Up to 120 words, sent to all of them at once. It's the third kind of writing in the series, after a story page and Instagram captions. A product page and a visit page come next.
I sent the same request to two fresh chats, word for word. I call them the first chat and the second chat. The six screenshots follow the test in order.
I ran both on 28 October 2026, in ChatGPT, in Think mode, in temporary chats with personalization off. It's one run, so I describe what happened here and no more.
The first chat
The first screenshot is the request. This chat had two files: the farm's record of dates, practices and terms, and what it may and may not say, compiled from public sources, not legal advice.
The second is its email. The facts are right: the roast days, shipping the day after, changes on the account page with no fee. In my view it's generic, even with the two files: any coffee brand could send "We appreciate your support", and this farm writes more simply.
The second chat
The third is the second chat's request. It is the same request, with the same two files. It also had two voice files: one on how the farm sounds, and one with lines the farm has approved, there to learn from and not to copy. So both chats were asked the same way. The files are the only thing I changed.
The fourth is its email. It's the shorter one, 81 words against 116, and the differences are in the lines themselves.
The subject line. The request asked for one.
The thanks. The request asked for thanks for staying.
The days. The request asked for the roast and ship days.
Who it's written to. The request said the email goes to every subscriber at once.
The request asked for a subject line, thanks for staying, the roast and ship days, and an email to every subscriber at once. How each was done is a small decision about how this farm writes to its subscribers, made once and written down. The second email has them, and they were in its files.
The second email slipped too, and each slip shows where the files need work:
A test is how a gap like that shows up.
The correction
The fifth screenshot is the one message I sent the second chat, as typed. It named three things: the email gives the facts in their own order and wording, "every 2, 4 or 6 weeks" is in figures, and "the day after we roast" reuses an approved line. Each time it sent the chat back to the part of its files to read again. It supplied no wording. The fix had to come from the files.
The sixth is the revised email. The first chat got no message and stays as it ran.
Fixed:
Partly fixed:
Two points fixed, one partly. That's all I'd claim.
Back to the question
Isn't that just a prompt?
The request was the same in both chats, so it isn't what separated the emails. The files were the only difference, and in this run the second email carries decisions that were in them. The request doesn't mention one thanks, the named days or a plain subject line. Those were in place before anyone asked. That's the difference between a request and a voice written down.
The second email isn't finished work, and that's what testing is for. A voice system for a real brand is built from the brand's own material and tested on its own writing before the brand relies on it. This is what that testing turns up, and each slip is one more thing to write down.
In my view, the plainness of the second email is what this farm sounds like: short sentences, the days named, nothing said about the reader that the farm can't back.
What would a new writer on your team have in front of them the next time your brand has to say thank you?