Brand Voice AI Prompts: Train ChatGPT on Your Guidelines

Paste your brand voice guide into ChatGPT and ask it to “write in our voice” and you get the same thing every time: competent copy that sounds like every other company whose guide says “confident, approachable, human.” The guide is not the problem. Guides are written for people, who fill in the gaps with taste. A model fills the gaps with the average of the internet.

My fix is to rewrite the guide as rules a model can check itself against, back the rules with a few real examples, and split the work into a setup prompt you write once and a check prompt you run on every draft. This post is that process, with the prompts.

Off-brand output was a problem before AI made it cheap. In Lucidpress’s 2019 survey of more than 200 organizations, 81% said they still dealt with off-brand content, and respondents who kept the brand consistent credited it with a revenue lift of up to 33%. The survey is old and self-reported, so treat the 33% as a direction rather than a promise. The 81% is the number I believe without hesitation, because every marketing team I have worked with recognises it.

Why “confident but approachable” fails as a prompt

A voice guide fails as a prompt when its instructions cannot be checked. “Confident” is a judgment. “No hedging words: maybe, perhaps, we think, might” is a rule, and a model can obey a rule. Most guides are 80% adjectives and 20% rules, and the adjectives are the part the model ignores, because it has no way to tell whether it has complied.

So the first job is translation. Every adjective in the guide has to become one of four things:

• A banned list. Words and constructions you never use. This is the highest-value rule type because it is binary.
• A required pattern. How you address the reader, what tense you use, how long a sentence runs, whether you use contractions.
• A claims rule. What the brand never says about itself or competitors, and what always needs a source.
• A structural rule. How copy opens, whether headings are questions or statements, whether you use lists.

A rule the model cannot verify from the text alone is not a rule yet. Keep translating until it is.

Step 1: extract the voice from work you already like

I start from output, not the guide. Pick five to eight pieces of copy your team agrees are the brand at its best, across at least two formats (an email and a landing page, say, or a post and a product description). Then run this:

You are a copy chief documenting a brand's writing style.
Below are [N] pieces of copy this brand considers its best work.

Read all of them, then describe the voice ONLY as checkable rules:
1. Words and phrases that appear repeatedly (quote them).
2. Words and constructions that never appear, even where most
   brands would use them.
3. Sentence length: typical range, and the longest sentence.
4. How the reader is addressed, and how the brand refers to itself.
5. How pieces open (first sentence pattern) and how they close.
6. Punctuation and formatting habits: contractions, lists, bold,
   exclamation marks, questions in headings.
7. Any claim pattern: what the copy asserts and how it backs it.

Do not use adjectives like friendly, confident, professional or
bold anywhere in your answer. If you cannot express something as
a rule, leave it out.

[paste the samples, separated by ---]

Why it works: banning adjectives forces the model to do the translation for you, and asking it to quote recurring phrases produces a vocabulary list you would never have written from memory. Read the output critically. It will overfit to the samples, so strike anything that is a quirk of one piece rather than a habit of the brand.

If you have a written guide, run it through the same prompt afterwards and ask the model to list every instruction in the guide it could not turn into a rule. That list is what your guide has been failing to communicate to humans, too.

Step 2: build the system prompt

The system prompt is the part you write once and reuse for months. It has four blocks. I keep mine under about 400 words. Longer system prompts get skimmed by the model in the same way long guides get skimmed by people.

ROLE
You write marketing copy for [brand], which [what it sells, to
whom, in one sentence]. You write as [we / the brand name / a
named person], never as "I" unless told otherwise.

VOICE RULES
- Address the reader as "you". Refer to the company as "we".
- Sentences average [N] words. No sentence over [N] words.
- Contractions: [always / never / in email only].
- Banned words and phrases: [paste list from step 1, plus:
  unlock, leverage, seamless, game-changing, empower, robust].
- Never open with a question, a definition, or "In today's".
- Never hedge a claim: no "may", "might", "could potentially".
  If a claim needs a hedge, cut the claim.
- Numbers: use digits, give the source in brackets, and never
  invent a figure. Write [NEEDS SOURCE] when you do not have one.

CLAIMS RULES
- We never say [best / #1 / leading] about ourselves.
- We never name a competitor in copy.
- We never promise results we cannot show a customer achieving.

FORMAT
- Headings in sentence case.
- Default to prose. Use a list only when the items are a list.
- Before returning copy, check every rule above and fix
  violations silently. Then return the copy only.

Why it works: the two rules doing the most work are the banned list and “write [NEEDS SOURCE]”. The first removes the vocabulary that makes AI copy sound like AI copy. The second stops the model from filling a persuasive gap with a statistic it made up, which is the failure that gets a brand into real trouble. The self-check instruction at the end matters more than it looks. Models comply better with rules they are told to verify.

Step 3: add three examples, not thirty

Rules describe the voice. Examples demonstrate it, and a model learns more from a demonstration than from a description - the few-shot technique that this step applies is worth understanding on its own. I add three examples after the system prompt, and make them different kinds:

1. A short piece in the voice, unedited. Your best 100 words.
2. A before and after. A paragraph of generic copy, then the same paragraph rewritten in the brand voice, with a one-line note on what changed.
3. A wrong example, labeled wrong. A paragraph that breaks three of your rules, with the rules it breaks named.

EXAMPLES

Example 1 (in voice):
[100 words]

Example 2 (generic, then rewritten):
Generic: [paragraph]
Rewritten: [paragraph]
What changed: [one line]

Example 3 (NOT in voice, for contrast):
[paragraph]
Breaks: [rule], [rule], [rule]

Why three: the before-and-after teaches the transformation, which is the actual job the model will do most of the time. The wrong example gives the rules teeth, because the model now has a concrete picture of the failure. Past five or six examples, output quality stops improving and the model starts copying structure and phrases from the samples wholesale, which produces copy that plagiarises your own past work.

Vary the formats across the three examples. Three email examples train an email voice and the model will write landing pages like emails.

Step 4: the voice-check prompt

The system prompt gets you an on-voice first draft most of the time. The check prompt is the one I run on every draft, including drafts humans wrote, and it is where the rules pay off, because now they are a checklist.

Check the copy below against the VOICE RULES and CLAIMS RULES
in the system prompt.

Return a table with one row per violation: the sentence, the
rule it breaks, and a rewrite. Then give a count of violations
by rule. Do not comment on anything that is not a rule
violation. Do not praise the copy.

If there are no violations, say "No violations" and nothing else.

[paste draft]

Why it works: “do not praise the copy” and “one row per violation” turn the model from a flattering editor into a linter. The count by rule is your diagnostic. If the same rule fails in every draft, the rule is either badly written or your writers disagree with it, and both are worth knowing.

Run the check on the model’s own output as well. The self-check instruction in the system prompt catches most violations but not all, and a second pass in a fresh conversation catches the rest, because the model is no longer defending its own draft.

Where to keep it

I keep the system prompt and examples somewhere every writer pulls from and only one person edits. A custom GPT, a Claude project, or a pinned document in your shared drive all work. The tool matters less than the ownership. A voice prompt with three editors drifts as fast as a voice guide with three editors.

Version it. When you change a rule, note the date and the reason. Six months in, the change log is the most useful document you have about how your voice actually works, because every entry records a case where the rules and the copy disagreed and someone had to decide.

For teams past a handful of writers, the same prompt belongs at the top of every other prompt in your library. My blog post prompts and email subject line prompts both assume a voice block is already in place, and they produce generic output without one. The voice block is also only one layer of the standing setup - the product facts, claims rules and output defaults around it are covered in the system prompt guide.

A worked example (hypothetical)

Imagine a B2B payroll software company whose guide says “expert, plain-spoken, reassuring”. Nothing in that sentence is checkable. Running their five best pieces through the step 1 prompt, the extraction comes back with rules like: sentences average fourteen words, the copy never uses “solution” or “streamline”, every piece opens with the reader’s situation rather than the product, the company is “we” and is never named mid-copy, and every number has a bracketed source.

Those rules become the system prompt. The wrong example in step 3 is a paragraph that opens with “In today’s fast-paced business environment”, calls the product a “seamless solution” and cites “studies show”. The first draft the model produces afterwards opens with “Your first payroll run with a new provider is the one that goes wrong”, which is a sentence the guide would have described as “reassuring” and the rules produced without needing the word.

The scenario is invented. The mechanics are not. The step where I see teams fail is step 1, because it is tedious, and the step they skip is step 4, because the first draft looked fine.

Common failures

The rules are adjectives in disguise. “Be concise” is not a rule. “No sentence over 25 words” is. Go back through the system prompt and test each line: could the model check this against the text with no judgment involved?

Too many examples of one format. Covered above, and the most common way a good voice prompt produces bad landing pages.

Long conversations drift. Twenty turns into a chat, the model weights the recent exchange over the system prompt and the voice slips. Start a fresh conversation for each piece and paste the system prompt again. Custom GPTs and projects do this for you.

The banned list gets stale. Models pick up new tics as they are updated. Reread a month of output every quarter, add the new offenders to the list, and expect the list to grow.

The guide and the prompt disagree. Once the prompt exists, the prompt is the guide. Retire the adjective version or rewrite it from the rules, so writers and the model are reading the same document.

FAQ

Does this work with Claude, Gemini and other models as well? Yes. Everything here is a plain-text instruction plus examples, and every major model handles that format. The banned list may need tuning per model, because each one has its own favourite filler words.

How long should the system prompt be? Under 400 words, plus the three examples. If you cannot fit the voice in that space, you have not finished translating adjectives into rules.

What about multiple voices, say a corporate brand and a founder’s personal voice? One system prompt per voice, with a shared claims block. Do not try to encode two voices in one prompt with a “switch to voice B” instruction. The model will blend them.

Can I skip the examples and rely on rules? You can, and the output will be rule-compliant and flat. The examples are where the rhythm comes from. Rules stop the model doing the wrong things. Examples show it the right ones.

This post is part of the prompt library for marketers. The voice block is the first thing to set up before running the social media caption prompts or the ad copy prompts, and it slots into the review stage of the AI content brief workflow. For the writer whose craft this systematizes, the career view is in AI for copywriters, and for the person who owns the voice across a whole team, the governance view is AI for brand managers. For where voice fits an AI marketing practice overall, see the complete guide to AI marketing.