To train AI on your brand voice, give it what you’d give a new copywriter: written voice principles, a vocabulary list, banned phrases and annotated examples of good, bad and edited copy. Package that into a custom GPT or shared project the whole team uses, then feed every human edit back into its instructions. The assistant doesn’t learn your voice once; the setup keeps teaching it.
Why “write in our tone” does not work
Most teams start by pasting a line like “write in a friendly, confident, professional tone” into a chat. The output comes back polished and interchangeable, because those three adjectives could describe almost any company. When instructions are vague, the model fills the gaps with its defaults: stock openers, lists of three, words like “seamless” and “unlock,” and a cheerful sign-off.
Three things are missing:
- Behavior, not adjectives. “Confident” means nothing to a model until you say what confident writing does and doesn’t do.
- Examples. Models imitate far better than they interpret. One annotated example teaches more than a paragraph of description.
- Consistency across people. When everyone writes their own prompt, output varies by who’s asking, and no correction carries over to the next person.
A note on “train”: most marketing teams don’t need fine-tuning, which changes the model itself and needs a large library of approved content. You’re giving a general model the right context on every request, which is cheaper and faster to change when positioning moves.
Documenting voice: principles, vocabulary and banned phrases
Write the voice guide for the model first and the brand book second. Every line should be something a reviewer could check a draft against.
Turn adjectives into rules
| Adjective in the brand book | Rule the model can follow |
|---|---|
| Confident | Make recommendations directly. Hedge only where there’s real uncertainty. |
| Friendly | Write to one reader as “you.” Use contractions. No exclamation points in body copy. |
| Expert | Include one concrete detail per paragraph: a number, a step or a named feature. |
| Plain-spoken | Keep sentences short. Use the buyer’s words, not internal jargon. |
| Not salesy | Say what the product does before saying why it’s good. No superlatives without proof. |
Aim for five to eight principles. Beyond that, they start to conflict.
Vocabulary
List the terms the model should use consistently: how your product name is written, the category you claim, what you call customers, feature names and the key phrases from your positioning statement. Add pairs where the team keeps slipping, such as “use ‘workspace,’ not ‘account’” or “say ‘clients,’ never ‘users.’”
Banned phrases
A banned list works better than a vibe. Include three groups:
- Generic AI tells: “unlock,” “elevate,” “game-changer,” “seamless,” “in today’s fast-paced world,” “let’s dive in.”
- Claims that need proof: “best,” “#1,” “guaranteed,” “industry-leading,” unless someone has signed off on the evidence.
- Brand-specific no-go words: competitor names in owned copy, retired product names, terms from a previous positioning.
Give a replacement where you can. “Instead of ‘seamless integration,’ name the integration and what syncs” is more useful than a bare ban.
Format rules by channel
Add a short block per format: length, structure, headline case, CTA style and emoji policy. Ads, lifecycle emails and LinkedIn posts need different rules, and the model won’t infer them.
Choosing examples: good, bad and edited
Examples do most of the work, so curate them. Uploading the whole blog teaches the model your average, including pages from before your last repositioning.
Use three kinds:
- Good examples. Three to five per format as a starting point, chosen by the person who owns that channel. Only include copy you’d publish today.
- Bad examples. Drafts that are grammatically fine but off-voice. A raw AI draft is often the best one, because it shows exactly the default you’re steering away from.
- Edited pairs. An AI draft next to the human-edited version, with a note on each change. These are the most valuable examples you’ll have, because they show the gap between the model’s default and your brand.
Label and annotate each one so the model knows what it’s looking at:
EXAMPLE: Onboarding email, day 3 (EDITED PAIR)
AI draft: "We're thrilled to have you on board! Let's unlock..."
Edited: "You've set up your first project. Here's the one setting..."
Why: Opens with what the reader did, not how we feel. Cut the
exclamation point and "unlock." One next step instead of three.Before the library goes live, check it:
- Every good example reflects current positioning and offers
- Every example is labeled good, bad or edited
- Every example has a one- or two-line annotation
- Each format the team produces has at least a few examples
- No customer data, unreleased plans or anything else your AI rules exclude
That last point matters. If you don’t have written rules for what can go into AI tools, set them first; an AI policy for marketing teams covers data, brand and review rules.
Building the assistant: custom GPTs, system prompts and shared instructions
Pick the tool your team already opens every day. An assistant nobody opens has no voice at all.
| Option | Best for | Where the voice lives | Watch for |
|---|---|---|---|
| ChatGPT custom GPT | Teams already working in ChatGPT | Instructions plus uploaded knowledge files | Files are searched, not always read in full |
| Claude Project | Teams already working in Claude | Project instructions plus project knowledge | Share it with the team, not just its creator |
| Gemini Gem | Google Workspace teams | Gem instructions | Check whether your plan supports files and sharing |
| API system prompt | High-volume formats like ad variants or product descriptions | A prompt inside an n8n workflow or internal tool | Someone must own and version the prompt |
Whichever you choose, structure the instructions in layers:
- Who we are and who we write for: one paragraph on the company, the buyer and the problem you solve.
- Voice rules: the principles table, written as instructions.
- Vocabulary and banned phrases.
- Format rules for each channel.
- Process rules: ask for the offer, audience and channel if they’re missing; flag any claim that needs proof; return two or three variants for short copy.
- Self-check: before answering, scan the draft against the banned list and the format rules.
Keep the core rules in the instructions field and put the example library in knowledge files where the tool supports them. Some tools cap instruction length, and uploaded files are often retrieved in pieces rather than read in full, so a rule buried on page 12 of a PDF may never be seen.
Start with one core assistant. Split by channel only when rules conflict, and keep one master voice document each assistant copies from, so a voice change is made once.
Use cases by team: ads, email, social and sales
The assistant writes from inputs, not from nothing. Each team needs a short brief template so requests arrive with the facts.
| Team | Typical tasks | Inputs it needs | Voice risk to watch |
|---|---|---|---|
| Paid ads | Headline and description variants, video hooks | Offer, audience, platform character limits, past winners | Clichéd hooks, unproven claims |
| Email and lifecycle | Subject lines, flow copy, newsletters | Segment, goal of the send, previous emails | Fake urgency, overly promotional tone |
| Social | LinkedIn posts, captions, repurposed content | Source material, the author’s personal voice notes | Sounding like every other post in the feed |
| Sales | Follow-ups, sequence steps, one-pagers | Call notes, deal stage, persona | Promising what marketing never claims |
Two notes. Founder and executive posts need a separate personal voice profile, since a person’s voice isn’t the company’s. And long-form SEO content needs more than voice: research, briefs and fact-checking. The AI content workflow for SEO covers that side.
The review and feedback loop
This is where most brand voice assistants fail. Editors fix drafts in the document, the fix never reaches the instructions, and the same mistake shows up next week.
Set up a loop instead:
- Every AI-assisted piece gets a human editor before it ships. The editor rates the edit as light, medium or heavy.
- Log corrections in one shared place: format, what the AI wrote, what changed and why. A simple form or spreadsheet is enough.
- Review the log on a fixed cadence, every two weeks at first. If the same fix shows up three or more times, it becomes a rule or a new edited pair.
- Version the instructions. Keep a short changelog so you can see what changed and roll back a rule that made things worse.
- Retest against a fixed set of briefs. Keep 10 to 15 standard requests covering each format and rerun them after every change to confirm the fix worked and nothing else broke.
The share of drafts needing heavy edits is your working metric. If it isn’t falling over a couple of months, the problem is usually the examples, not the rules.
Two failures to watch: one senior person’s taste quietly becoming the house style, and rules piling up until they contradict. The fix for both is a single owner who approves every change, while anyone can submit corrections.
Keeping it current as positioning changes
A voice assistant built on last quarter’s messaging will confidently write last quarter’s messaging. Update it when any of these change:
- Positioning, category or core message
- Product names, pricing or packaging
- Ideal customer profile or target segments
- Legal or compliance rules on claims
- Channels, such as adding SMS or a new social platform
Even without a trigger, review it quarterly. Remove examples that no longer reflect the brand; stale examples are worse than none, because the model copies them. Update the vocabulary and banned list, then rerun the test briefs.
Building this kind of shared assistant, with the workflows and review loop around it, is part of my AI automation work. The goal is fewer rewrites, not more content.
Get it built
If your team’s AI drafts all sound alike, and not like you, I can build the voice guide, example library and shared assistant, then set up the loop that keeps them current. Start with a Growth Audit, $1,500 fixed and credited if we continue, or add AI automation from $2,500/mo. See pricing or get in touch.