How to Train AI on Your Brand Voice (Free Template)
Most brand-voice prompts fail because they describe your voice instead of showing it. Here's a reusable spec that fixes that.
An LLM cannot infer "professional yet approachable." It can only copy patterns it can see. Stop describing your voice and start showing it.
To train AI on your brand voice, give the model three things it can actually copy: 5–10 real samples of your best writing, a set of hard rules (words to use and ban), and 3–4 before/after rewrites that show your voice correcting a generic draft. Adjectives like "professional yet approachable" do almost nothing. Patterns, examples, and explicit rules do almost everything. This guide gives you a copy-paste template to do it in under an hour.
If you have ever pasted your brand guidelines into ChatGPT and still gotten back something that sounds like a LinkedIn motivational post written by a committee, you are not doing it wrong — you are doing what everyone does, which is the wrong thing. The problem is not the model. The problem is that you are describing your voice when the model can only learn from examples of it.
This is the single most important idea in this article, so it is worth stating plainly: a large language model does not understand adjectives the way a creative director does. When you write "our tone is bold but warm, confident but never arrogant," the model maps those words onto the internet-average of "bold," "warm," and "confident" — which is exactly the bland, over-polished register you were trying to escape. To get your voice, the model needs to see your voice.
Why "add your brand guidelines to the prompt" doesn't work
Traditional brand guidelines are written for humans. They lean on abstraction — values, personality archetypes, mood boards — because a human reader fills in the gaps with judgment. An AI model has no judgment to fill the gaps with. It has pattern-matching.
Consider the instruction "be conversational." A junior copywriter knows that means short sentences, contractions, the occasional fragment, and a question now and then. The model knows "conversational" appears near roughly ten million documents of varying quality, and it averages them. The result regresses to the mean — which is why AI content so reliably sounds like AI.
The fix is to replace abstraction with evidence. Instead of telling the model who you are, you show it what you have already written and give it rules specific enough that there is nothing left to average.
The Brand Voice Spec: five components AI can actually use
A "Voice Spec" is a single document, written for a machine, that contains everything the model needs to reproduce your voice. It has five parts, in rough order of importance.
1. Samples (the most important part)
Gather 5–10 pieces of your best, most on-voice writing — the ones you would frame on a wall. Aim for 1,500–3,000 words total for short-form work, more if you write long-form. Use real published pieces, not aspirational ones; the model copies what you give it, so a sample that is 80% right will make your output 80% right.
Pick samples that are varied (an intro, a product description, a hard-news update, a cheeky social caption) so the model learns your range, not just one register.
2. Hard rules: words to use, words to ban
This is the highest-leverage, lowest-effort component. List the words and constructions you always use and the ones you never use. Be ruthless and specific:
- Always: contractions; "you" and "we"; sentences under 25 words; active voice.
- Never: "elevate," "unlock," "leverage," "in today's fast-paced world," "the power of," em-dash pile-ups, exclamation points in body copy.
- Spelling/format: "ecommerce" not "e-commerce"; sentence case headlines; Oxford comma on.
A ban list does something adjectives can't: it removes the specific tics that scream "AI wrote this." Ten banned phrases will do more for your voice than ten paragraphs of personality description.
3. Before/after rewrites
This is the secret weapon almost no one includes. Show the model 3–4 pairs: a generic sentence, then the same idea in your voice.
Before: "Our platform empowers businesses to unlock their full marketing potential."
After: "Ivon runs your marketing while you run your company."
Before/after pairs teach the transformation — the move from default register to your register. The model learns not just the destination but the direction of travel, which generalizes far better than samples alone.
4. Audience and context
One tight paragraph: who you are talking to, what they already know, what they care about, and what action you want. "We write for time-poor founders of 5–50 person companies who are skeptical of marketing hype and want proof, not adjectives." This anchors register — the same brand sounds different talking to a CFO than to a teenager.
5. A few adjectives — last, not first
Now, and only now, add a short personality line. Three to five traits, each paired with a concrete instruction so the word has somewhere to land: "Confident — make claims directly and don't hedge with 'we believe' or 'arguably.'" Adjectives are the seasoning, not the meal.
How to train AI on your brand voice in five steps
Here is the full workflow, start to finish. It takes about an hour the first time and minutes thereafter.
- Collect samples. Drop 5–10 of your best pieces into one document. This is 80% of the work and 80% of the result.
- Write your hard rules. Spend 15 minutes building your always/never/format lists. Mine your own published work for the patterns.
- Build before/after pairs. Take three generic sentences and rewrite them in voice. Keep the contrast sharp.
- Assemble the Voice Spec using the template below, and save it as a reusable file — a system prompt, a custom GPT, a "brand voice" setting, or a project knowledge file, depending on your tool.
- Test, correct, and version. Generate three pieces, mark every off-voice moment, and feed the corrections back as new rules or new before/after pairs. Your spec should get sharper every month, not sit frozen in a folder.
Copy-paste Voice Spec template
# BRAND VOICE SPEC — [Brand]
## Audience & context
We write for [who], who already know [X] and care about [Y].
Every piece should move them toward [action].
## Voice samples
[Paste 5–10 of your best on-voice pieces here, labeled by type.]
## Hard rules
ALWAYS: [contractions, second person, sentences < 25 words, ...]
NEVER: [list of banned words/phrases/tics]
FORMAT: [spelling, capitalization, punctuation conventions]
## Before / after
1. Before: "[generic]" → After: "[in voice]"
2. Before: "[generic]" → After: "[in voice]"
3. Before: "[generic]" → After: "[in voice]"
## Personality (seasoning, not the meal)
- [Trait] — [concrete instruction]
- [Trait] — [concrete instruction]
## Instruction
Write the following in this exact voice. Obey the hard rules.
Match the samples and before/after direction. Do not invent a new tone.
Tips and reminders
- Refresh quarterly. Add your newest best work and retire stale samples. A voice spec is a living document.
- One spec per channel if needed. Your blog voice and your X voice may differ; clone the spec and swap the samples and rules.
- Don't over-prompt mid-draft. If output drifts, fix the spec, not the individual generation. One-off corrections don't compound; spec changes do.
- Keep a human in the loop. The spec gets you to 90%. A final read catches the 10% that builds trust — and feeds tomorrow's rules.
- The spec scales beyond you. The same document that trains ChatGPT can onboard a freelancer, a new hire, or an autonomous marketing system.
Where this is heading
Pasting a spec into a chat window every time is the manual version. The direction of travel is systems that hold your brand voice permanently — that absorb your samples, rules, and corrections once and then apply them across every blog post, ad, and caption without re-prompting. That is the premise behind Ivon's autonomous marketing platform: deep brand integration so the voice is the default, not a setting you re-enter. If you are a small team trying to publish more without sounding generic, a durable voice spec is the foundation — and scaling content with AI only works once that foundation is in place.
Get the spec right and the model stops guessing. It stops averaging. It starts sounding like you — because, for the first time, you have shown it what you actually sound like.
Frequently asked questions
How many writing samples do I need to train AI on my brand voice?
For short-form content, 5–10 strong samples totaling 1,500–3,000 words is enough. For long-form like blogs, aim higher — 10,000 words or more. Quality matters more than quantity: pick your best, most on-voice pieces, because the model copies what you give it. A handful of excellent samples beats a pile of mediocre ones.
Why does AI ignore my brand guidelines?
Because traditional guidelines describe your voice in adjectives ("bold," "warm," "conversational"), and a model can't infer behavior from abstraction. It maps those words onto internet-average usage and regresses to the mean. Replace descriptions with evidence: real samples, explicit always/never rules, and before/after rewrites the model can copy directly.
Can I just put my brand voice in a custom GPT or system prompt?
Yes — that is the ideal place for your Voice Spec. A custom GPT, system prompt, project knowledge file, or "brand voice" setting all work as a permanent home for the spec, so you don't re-paste it every time. The contents matter more than the container: samples, hard rules, and before/after pairs are what actually move output.
What's the difference between brand voice and brand tone?
Voice is your consistent personality — it stays the same everywhere. Tone is how that voice flexes by context: more upbeat in a launch announcement, more measured in a support reply. Your Voice Spec captures the constant voice; the audience-and-context section tells the model how to flex tone for the situation.
How often should I update my AI brand voice spec?
Refresh it quarterly, and whenever you spot a recurring off-voice pattern. Add your newest best work, retire stale samples, and convert every correction you make into a new hard rule or before/after pair. A spec that improves monthly compounds; one frozen in a folder slowly drifts out of date.