How to train AI on your brand voice

"Write in a friendly, professional tone" is the instruction nearly everyone starts with, and it's nearly useless. Almost every business thinks it's friendly and professional, so the AI has nothing to work with.

Here's a method that does work. It takes about half an hour, and you only need to do it once.

Step 1: Collect your best writing

Pick 5 to 10 pieces that sound the way you want to sound. Use real things you've published or sent: posts, emails, a page on your site.

Be choosy, because the AI will learn your habits, good and bad.

• Choose work you'd happily write again. Skip anything written by committee.
• Keep to one kind of content. Your LinkedIn voice and your contract voice are different. Teach them separately.
• Remove what you don't want repeated, such as old offers, or a phrase you've gone off.

If you don't have much, write one short piece by hand, as if to a friend. That's often enough to begin with.

Step 2: Describe it in concrete terms

Now read your samples and write down what you actually see. Avoid adjectives like "engaging". Look for things that can be checked:

Vague	Concrete
Friendly	Uses "we" and "you". Short sentences. Contractions always.
Professional	No slang. No exclamation marks. Numbers over adjectives.
Funny	One dry aside per piece, usually at our own expense.
Confident	States opinions without "I think". Doesn't hedge.

Four questions help here:

1. How long are your sentences, and do they vary?
2. Which words do you use often that others in your field don't?
3. How do you usually open? How do you usually end?
4. What would you never say?

The last one is the most useful. "We never say 'solutions', 'passionate' or 'elevate'" removes a surprising amount of AI flavor straight away.

Step 3: Separate voice from facts

Voice is how you say things. Facts are what you say. Keep them apart. Your samples teach voice, and your facts belong in reference material. If you mix them, the AI may repeat an old price because it appeared in a sample.

Step 4: Test it on something real

Don't test with "write a post about our business". Use a request you really need, then compare the draft to your samples:

• The read-aloud test. Does it sound like something you'd say? Where do you stumble?
• The swap test. Could a competitor publish this by changing the name? If so, it needs more of your facts.
• The ending test. Does it stop where you'd stop, or add a tidy conclusion?

Step 5: Turn every correction into a rule

When you fix a draft, ask why. If you changed "We're thrilled to" into "We're", the rule is "Never say we're thrilled or excited". After a few rounds of this, you'll be making far fewer corrections.

Good rules are specific:

• Never open with a question.
• One idea per post.
• End on a practical detail, not a call to action.
• Say what it costs, if the price is known.

Common mistakes

Too many samples of mixed quality. More isn't better if half of it is weak. Five great pieces beat 50 average ones.

One voice for everything. A press release and a Facebook post shouldn't sound the same, even from one brand. Your values stay, but formality and energy shift. Build a separate voice for each.

Copying a famous writer's style. You get a parody of them, not a voice of your own.

Never reviewing it. Voices drift as a business changes. Look again every few months.

How this works in DeepRit

In DeepRit, each writer has its own voice settings for tone, formality, warmth and energy, plus rules for what to do and what to avoid. Paste a writing sample and it's analysed once into a short style note. Only that note is sent with later requests, which keeps costs low. Then use the writer's Test step on a real request before you save anything.

Next, give it something to say: how to give AI your business data.