Your AI Knows Your Writing Style. Does It Actually Know Your Brand?

Ask an AI to write your product update "in a friendly, confident tone" and you'll get something that reads fine. Warm opener, short paragraphs, maybe a light joke.

Then you read it properly. It calls your customers "users" when your whole sales deck says "practice managers". It promises "enterprise-grade security" you have never had audited. It describes a feature you retired in the spring, using the old name.

The AI nailed your tone and missed your brand. A writing prompt tells an AI how to sound. An operational brand guide tells it who you serve, what you sell, which words you use, and which claims you must never make. If you produce content with AI every week, as a founder, a marketer or an agency, you need the second one. And every tool you use needs to read the same copy of it.

What a writing prompt gives your AI

Most "brand voice" prompts I see are a row of adjectives. Friendly, professional, confident, no jargon. Some add a sample paragraph and the line "write like this".

Models are good at this part. They have read millions of friendly, confident paragraphs, so matching a tone costs them nothing. That's why AI copy can feel on-brand at a glance.

Adjectives carry no facts, though. "Friendly" doesn't tell the model that your customers run dental practices and have been burned by software contracts before. "Confident" doesn't tell it the difference between confident and making things up.

The prompt also lives in the wrong place. Your marketer keeps one in ChatGPT's custom instructions. Your agency keeps another in a Claude project. Your developer has a third buried in Cursor rules. Six months later you have three brands, each drifting a little further from the one on your website.

What your brand contains beyond tone

Think of tone as the accent. Your brand is everything you'd brief a new hire on before letting them email a customer. Four things matter most when that new hire is an AI.

Your target customers

The AI needs to know who is reading, what they already know, and what they worry about. "SMEs" is a category. "A two-person e-commerce team that has outgrown Shopify's blog and has no developer" is a reader. The second description changes every sentence the AI writes: which problems it leads with, which words it explains, how much it assumes.

Your products, as they are today

Models fill gaps with plausible guesses. If your guide doesn't list your current plans, features and names, the AI will invent a sensible-sounding version. It might describe a free trial you stopped offering, or merge two products into one. Write down what each product does in a sentence, what it doesn't do, and any names you've retired.

Your terminology

Every business has words it says and words it avoids. In Garchi CMS, a project is a "space". If an AI writes "workspace" in a help article, a customer goes hunting for a menu that doesn't exist. Your words might be "clients" over "customers", "members" over "users", or a product name with odd capitalisation that everyone gets wrong.

The claims you must never make

This is the part a tone prompt can't cover, and it's where AI content gets expensive. An AI trying to sound confident will happily write "fully GDPR compliant", "guaranteed results", "the UK's leading" or "clinically proven". In the UK, the ASA expects you to hold evidence for claims like these before you publish them. A physio clinic that lets an AI write "cures back pain" on its website has a problem no friendly tone will fix.

So list them. Claims you never make, and claims you only make with a source attached.

Writing prompt vs operational brand guide

A writing prompt is a style note. An operational brand guide is a working document your whole team, and every AI it uses, writes from.

You still want the tone. It just becomes one section of a bigger document instead of the whole thing.

Keep one guide where every AI tool can read it

A brand guide in a Google Doc helps nobody if the AI never opens it. Store the guide next to the content instead, and any tool that writes for you can find the rules in the same place it finds your pages.

That’s why we built brand voice into Garchi CMS at the space level. A space holds one brand’s content: its pages, blog posts, products and social posts. Each space can now carry its own brand voice.

Setup takes two steps. Answer a short questionnaire on your audience, values, tone, formality, spelling, words to avoid, examples, social style and house conventions. Garchi turns your answers into a first draft in Markdown, up to 20,000 characters. Edit it as much as you like, or paste in a guide you already have.

Any AI agent connected to that space through Garchi’s MCP server can read the guide, whether that’s Claude, Cursor or another MCP tool. Tell your agent to check the brand voice before it drafts. Your marketer, your agency and your developer can drop their three private prompts and work from one guide. Edit it once and the next draft uses the new version.

Agencies get the most from this. Give each client its own space and each one gets its own guide, so the AI writing for the dental group won’t borrow phrases from the fintech client you worked on that morning.

Two honest limits. The guide shapes how an AI writes. It doesn’t fact-check what it writes, and an agent only follows it if it reads it. Reading the guide also gives an agent no extra permissions. Pages and social posts land in Garchi as drafts, and a person on your team approves them before they go live. Blog posts and products work the same way unless you schedule them to publish. Brand voice is part of the Garchi Pro and Business plans.

How to write an operational brand guide in an afternoon

You don't need a brand agency for this. You need the stuff that's already in your head, your sales calls and your support inbox, written down.

1. Describe your customer in their own words. Their job, what they're trying to get done, and what they've been burned by before. Lift phrases from real emails and calls.
2. List your products as they stand today. One line on what each does, one line on what it doesn't, and any old names the AI should never use.
3. Build a glossary of "we say" and "we don't say". Ten to twenty entries covers most of it.
4. Write a claims register. Claims you never make, and claims you only make with a link to the evidence.
5. Add examples. Two pieces of copy you love and one you hate, with a sentence on why for each.
6. Note your conventions: British or American spelling, how you capitalise product names, how you sign off, what a call to action looks like.
7. Put a name and a review date at the top. A guide nobody owns goes stale in a quarter.

The claims register is the section most teams skip, so here's what a small one can look like:

## Claims
Never say:

"Guaranteed" anything
"The UK's leading..." (we have no ranking to back it)
"Fully GDPR compliant" (say "built with GDPR in mind")

Only say with a source:

Customer numbers: use the figure on the pricing page
Time saved: only from a named case study

A five-minute test

Open the AI tool you use most for content. Start a fresh chat with your usual brand prompt and ask it three questions: who are our customers, what are our products called, and what must you never claim about us?

If it answers all three correctly, your setup is in good shape. If it guesses, remember that it has been guessing the same way in every blog post, product page and LinkedIn caption it has written for you so far.