A client came to me last year with a problem she couldn’t quite name. She’d been using ChatGPT for three months. The content was fine. Technically correct, grammatically sound, going out on schedule.
Then one of her regulars messaged to ask whether she’d changed her social media person.
She hadn’t changed anything. She’d handed her brand voice to a tool that had never been told what her brand voice was. Her customers noticed before she did, which is the part I’d take seriously if I were you. The gap is visible from outside the business long before it’s visible from inside it.
Why does AI content sound generic?
Because generic is what got asked for.
Most people type a short prompt and expect the tool to work out the rest. Write a social media post for my jewellery business. That’s close to a real request I’ve been shown, and the result sounded like every other jewellery business on the internet, because nothing in that sentence describes this jewellery business.
The model isn’t being lazy. It’s doing exactly what was asked. Given nothing specific about who you are, it falls back on the average of everything it has ever read, which is a long way of saying it writes for everyone.
Will a better prompt fix it?
No, and this is where most of the effort gets wasted.
People spend hours rewriting the prompt, hoping one perfect instruction will solve the voice problem in a single go. It won’t, because a great prompt sitting on top of no context is still generic. The prompt is only the question. The context is everything the tool knows before you ask.
Anthropic put this plainly in a piece on context engineering published in September 2025. Building with language models, they write, is becoming less about finding the right words and phrases for your prompts, and more about what configuration of context is most likely to produce the behaviour you want.
So the businesses getting genuinely good output aren’t the ones with the best prompts. They’re the ones who gave the AI the right context, and not much else.
What should you actually write down?
Four things. Not a folder, and not a shared drive full of documents nobody opens. Four.
- How you sound. Real examples of your own writing, plus the words you never use. That second half does far more work than people expect.
- Who you’re talking to. Not demographics. The actual words your customers use when they describe the problem, the worry they have but rarely say out loud, and the thing they’re really after when they ask for something else.
- The facts it must never guess. Your prices, your claims, your product names, and the two things you’re not allowed to say. This is missing from nearly every brand voice document I get handed, and it’s the one that stops the tool inventing something you then have to correct in public.
- What good looks like. One or two pieces you were genuinely happy with. Anthropic’s advice to their own engineers is to curate a small set of diverse, canonical examples rather than a long list of edge cases, on the basis that for a language model, examples are the pictures worth a thousand words. For a brand that holds exactly.
Notice what isn’t on that list. Your founding story, your values statement, the full service menu, three years of newsletters. All of it interesting to a human reader. None of it helping the tool sound like you.
How specific should a brand voice document be?
Specific enough to rule things out. Loose enough that you’ll actually use it.
Anthropic call this the right altitude, and they name two ways of getting it wrong. At one end, people write brittle rules trying to cover every possible case, which produces something fragile that gets harder to maintain over time. At the other end, people write vague high-level guidance that gives no concrete signal at all.
Both failures turn up in brand voice documents constantly, and once you can see them you can’t unsee them.
The vague end is a list of adjectives. Friendly, professional, authentic. Almost every business in the country would tick all three, which means those words rule nothing out and change nothing about what comes back.
The brittle end is the forty-page brand bible with a rule for every scenario anyone could imagine. Nobody reads it, including the AI, because by the time you’ve pasted it in you’ve spent most of the available attention on instructions rather than on the job in front of it.
Two paragraphs of your own writing beat any adjective you could pick.
The version that works sits between the two, and it’s shorter than most people expect. Three or four sentences on how you open a piece. A list of the words you never use. Two paragraphs of your own writing you’d be happy to see again.
Is more of it better?
It isn’t, and that’s the half I got wrong when I first wrote about this in May.
What I said then was load the brand voice document, load the audience intelligence, load the business context. Reasonable enough, except it reads as an instruction to pour everything in, and that’s exactly where it stops working. Context is finite. Anthropic describe an attention budget that every added token spends a bit of, alongside a documented effect called context rot, where the model’s recall gets less reliable as the window fills up.
So the goal was never the most context. It’s the smallest set of high-signal things that gets you the outcome you want. The longer version of that argument, with the sources, is in why context beats prompts every time.
One caution Anthropic add, which I’d repeat to any business owner about to delete half their notes. Minimal doesn’t mean short. You still have to give it enough to work with.
How do you stop retyping it every time?
You package it once instead of rebuilding it from memory every morning.
Write the four things down, keep them somewhere reusable, and load them as a set. That’s what a skill is, and it’s the rung of the ladder almost nobody is selling. I wrote about the difference between a prompt, a skill and an agent in prompt, skill or agent.
For a New Zealand small business this is easier than it is for anyone large, which isn’t the usual direction of travel. You already know how you sound, who you sell to, and what you’re not allowed to claim. Writing that down is an afternoon. A big organisation needs six people in a room to agree on it first.
If your AI content is embarrassing you or exhausting you, it isn’t the tool. It’s that nobody has told the tool who you are.
Which of those four would take you longest to write down? That’s usually the one worth doing first. If you want a hand working out what to keep and what to cut, let’s talk.
Sources
- Anthropic, Effective context engineering for AI agents, 29 September 2025. Source of the context engineering framing, the right altitude and its two failure modes, the canonical examples guidance, the attention budget, and the caution that minimal does not mean short.
- Chroma, Context Rot, cited by Anthropic in the piece above. Source of the finding that recall accuracy drops as the context window fills.
- The client story is my own work, told with her details removed.
