Samples of your real writing
Twenty pieces you would happily publish again. Emails, posts, product pages. This teaches rhythm and vocabulary better than any description could.
How to Make AI Write in Your Brand Voice, Not Its Own
How-to — — by Mahmoud Zalt
Make AI write in your brand voice using samples of your real writing, a short tone brief, and a do-not-say list it checks every time.
You can usually tell within one sentence. The em dashes, the tidy rule of three, the word "journey" where nobody needed a journey. It reads fine, and it does not read like you at all.
The reason is boring and fixable. A general model has read a very large amount of average marketing copy, so average marketing copy is its default setting. It is not failing at your voice. Nobody has shown it your voice yet.
In Sistava, voice lives across the same layers as everything else you customize. The tone brief goes in the short written page your AI Employee carries into every task. Your real writing goes into trained knowledge, so it can look at how you actually phrase things. The do-not-say list becomes a standing rule. And every time you correct a line, that correction sticks in memory so you never have to make it twice.
At a Glance
- 20
- Samples of your own writing, the single biggest lever
- 1 page
- The tone brief that travels with every task
- 15 words
- A do-not-say list is usually about this long
- Once
- How many times you should have to fix the same word
What actually teaches AI your brand voice?
Examples teach voice far better than adjectives do. Telling AI to be "warm but professional" barely moves the output, because those words mean something different to every writer alive. Showing it twenty things you actually published moves it a lot, because rhythm, sentence length, and word choice are all sitting right there in the samples.
So the order matters. Collect the writing first, then write the brief to explain what the samples cannot show, such as when to be blunt and when to soften. Adjectives on their own produce a polite average. Adjectives on top of real samples produce something recognisable.
Benefits
A short tone brief
Half a page naming how you sound, who you are talking to, and when to change register. It travels with every single task.
A do-not-say list
The words, phrases, and punctuation you never want to see. This is the fastest cleanup you will ever get, because it removes the tells.
Your corrections, said out loud
When a line is wrong, say why. Silent rewrites teach nothing. Spoken corrections become permanent preferences.
How do you write a tone brief that works?
A tone brief works when it makes decisions rather than describing a mood. "Friendly and approachable" is a mood and it changes nothing. "Start with the answer, never open with a question, use contractions, never use exclamation marks" is a set of decisions, and every one of them shows up in the next draft.
The most useful trick is a pair of contrasts. Write five lines the way you would say them, then write the same five lines the way a generic brand would. Put both in the brief under the headings "sounds like us" and "does not sound like us". Contrast is easier to copy than description.
Then add the register rules. Most businesses have at least two voices: one for a happy customer and one for an angry one. Say which is which. A brief that covers only the good day will produce something tone deaf on the bad day, and the bad day is when tone matters most.
Why does a do-not-say list matter so much?
Because most of what makes AI writing recognisable is a small set of repeat offenders. Ban fifteen words and the output stops sounding synthetic almost immediately. It is the highest return per minute of anything in this article, and it takes about ten minutes to write.
Build the list from your own irritation. Read three pieces of AI output, circle every word that made you wince, and that is your first draft. Add the industry clichés you have banned in your own team, plus any punctuation you dislike. Keep adding to it for a month and then it stabilises.
- Words you never use: pick your own, but most lists start with "unlock", "elevate", "delve", and "journey".
- Structures you hate: the rhetorical question opener, the three-item list in every paragraph, the closing summary nobody asked for.
- Punctuation rules: whether you allow em dashes, exclamation marks, or semicolons at all.
- Claims you cannot make: anything about results, timelines, or guarantees that your legal or regulatory position does not allow.
- Names and spellings: how your product is capitalised, and what you call your customers.
What does this look like for a real brand?
Nadia runs a small skincare brand out of Rotterdam. Nine thousand email subscribers, about forty products, and a voice she describes as dry and unimpressed, because her whole positioning is a reaction against wellness language. She writes everything herself and it was eating five hours a week.
Her setup took about an hour. She pasted in twenty-two of her own newsletters, the ones with the highest reply rates, plus eight product descriptions she liked. Her tone brief was six lines long: short sentences, no more than two per paragraph, never open with a question, assume the reader is smart, name the ingredient before the benefit, and never sell an emotion.
Her do-not-say list had fourteen entries. "Glow up", "self care ritual", "clean beauty", "pamper", "indulge", "treat yourself", "radiant", "transform", and a rule against exclamation marks. That list alone changed the drafts more than anything else she did.
Two months later she still writes the opening paragraph of the weekly newsletter herself, because that is the part her readers reply to, and the AI Employee drafts the rest. Her five hours became about ninety minutes. Forty product descriptions that had been on her list for a year got written in one afternoon, and she edited roughly one line in three.
| What Nadia gave it | Where it lives | What it fixed |
|---|---|---|
| 22 newsletters, 8 product pages | Trained knowledge | Rhythm, sentence length, her actual vocabulary |
| Six lines of tone rules | The brief, applied every task | Structure and where the benefit sits |
| 14 banned words and phrases | A standing rule | The wellness clichés she built the brand against |
| "That is too warm, cut the last line" | Memory, after one correction | The sign-off she kept deleting by hand |
| Nothing | Her opening paragraph | She still writes it, on purpose |
What can a brand voice setup never fix?
It cannot invent a voice you do not have. If your existing writing is inconsistent, the AI Employee will copy the inconsistency faithfully, because that is what the samples show. In that case the real work is deciding how you want to sound, and the setup just makes the decision stick.
It also cannot supply lived experience. The line about the customer who called you at eleven at night is yours and nothing can generate it. Expect strong drafts, not finished pieces, on anything that depends on a story only you were present for.
- It will not write your personal anecdotes, because it was not there.
- It will copy your bad habits along with your good ones if both are in the samples.
- It cannot judge whether a joke is too far for a specific customer on a bad day.
- It will drift on very long pieces, so voice holds better across many short drafts than one enormous one.
- It will not fix a message that is wrong, only the way that message sounds.
How to teach your voice in one sitting
- Collect twenty samples you would publish again — Emails, posts, landing pages, replies to customers. Favour the ones that got the best response over the ones you liked writing.
- Write the do-not-say list first — Fifteen words, phrases, and punctuation habits you never want to see. Ten minutes, and it changes the output more than anything else on this list.
- Write six lines of tone rules — Decisions, not moods. Sentence length, where the benefit goes, how you open, how you close, what you never claim.
- Add a sounds-like and does-not-sound-like pair — Five lines written your way, and the same five written the generic way. Contrast teaches faster than description.
- Correct out loud for two weeks — Say what was off and why, once. Do not silently rewrite. Every spoken correction becomes a permanent preference.
Comparison
| Dimension | Traditional | With Sista |
|---|---|---|
| Openings | Starts with a rhetorical question or a big claim. | Starts where you always start, with the point. |
| Vocabulary | Reaches for the industry cliché every time. | Uses your words, and avoids your banned ones. |
| Rhythm | Even, tidy, slightly airless paragraphs. | Your actual sentence lengths, copied from real work. |
| Bad news | Over-apologises or sounds corporate. | Follows your register rule for the hard emails. |
| Your editing | You rewrite most lines. | You edit roughly one line in three. |
Frequently asked questions
FAQ
How many writing samples does AI need to learn a brand voice?
About twenty is the point where it stops guessing and starts copying. Five is too few to show a pattern and a hundred adds very little over twenty. Pick pieces that got a real response, and mix formats so it sees how you sound in an email as well as on a page.
Why does AI writing still sound generic after I describe my tone?
Because descriptions like "warm and professional" mean something different to everyone, so they barely change the output. Swap adjectives for decisions such as sentence length, how you open, and what you never claim, then add real samples. Description plus examples works, description alone does not.
What should go on a do-not-say list?
The words that make you wince, the sentence structures you hate, punctuation you avoid, claims you are not allowed to make, and the correct spelling of your own product names. Around fifteen entries is normal. It is the fastest single improvement you can make to AI copy.
Can one AI Employee handle different voices for different channels?
Yes, if you write the register rules into the brief. Say plainly how a support reply differs from a newsletter and from a sales follow-up. If two channels are genuinely far apart, it is cleaner to give each one its own role rather than asking a single brief to do the splits.
Will readers be able to tell an AI Employee wrote it?
Less than you fear, once the do-not-say list and your samples are in place, because the tells are mostly vocabulary and structure. What still gives it away is missing specifics: names, numbers, and stories only you know. Keep supplying those and the writing reads as yours.
Do I need to rewrite my brand guidelines to do this?
No. Existing guidelines help if you have them, but a six line tone brief plus twenty real samples outperforms a forty page guideline document that nobody follows. If you have the long document, upload it as knowledge and keep the brief short.
Voice is the part people assume is hardest and it is usually the fastest win. You already have the samples, you already know which words you hate, and the rest is one sitting of writing things down that were previously only in your head.
If your problem is facts rather than tone, the companion piece on training AI on your own company documents is the one to read next. If it is rules and approvals, start there instead.