How to Create a Brand Voice Guide for AI Writing (With a Reusable Prompt)
Most teams using ChatGPT or Claude to draft content hit the same wall in the first month: the writing is fine, but it doesn't sound like them. Every writer prompts differently, the model falls back on its default polished-and-generic style, and your brand slowly turns into a patchwork.
The fix is a brand voice guide for AI writing: one page, specific enough that a model can follow it and a human can score against it. Below is a complete template you can copy, the system prompt that turns it into a reusable instruction, and a three-post test to check for drift.
Quick answer: A brand voice guide for AI writing is a one-page document with five voice traits, do/don't sentence pairs, a banned-phrase list and two sample paragraphs. Paste it into a system prompt (ChatGPT custom instructions or a Claude Project), then run three test posts and score each against the guide to catch drift.
Why AI needs a voice guide (and "friendly and professional" isn't one)
Left alone, a language model writes in a safe, neutral register that could belong to any company in your category. Adjectives like "friendly" or "professional" don't change that much, because they mean something different to every reader. Models respond far better to concrete material: examples, explicit rules and phrases to avoid.
That is what this guide is built from. If you've already worked on getting a model to sound less robotic, this is the document that makes those fixes repeatable across a whole team. It also helps with the cleanup stage: a clear standard makes editing AI drafts faster, because you're checking against a rule instead of arguing about taste.
Step 1: Pick five voice traits
Choose five traits, no more. Each trait gets a short definition and a "but not" boundary, because the boundary is what stops a trait from sliding into its cheap version. Pull them from your best existing content: read 10 pieces you're proud of and note what they have in common.
Here is an example for a fictional analytics company, Brightside:
| Trait | Means | But not |
|---|---|---|
| Plain | Short words, one idea per sentence | Dumbed down or vague |
| Direct | Lead with the answer or the point | Blunt or dismissive |
| Practical | Every section ends in something the reader can do | A list of tips with no context |
| Warm | Second person, contractions, a light touch of humor | Chatty, jokey or full of exclamation marks |
| Honest | Say what we don't know; admit trade-offs | Hedging every claim |
Step 2: Write do/don't pairs
One before-and-after sentence per trait is the most useful part of the guide for a model, because it shows the transformation instead of describing it. Write the "don't" the way the AI would actually write it by default.
| Trait | Don't | Do |
|---|---|---|
| Plain | "We leverage data-driven insights to optimize outcomes." | "We use your numbers to show what's working." |
| Direct | "In today's fast-paced landscape, many marketers wonder about reporting." | "Most weekly reports are too long. Here is a shorter one." |
| Practical | "Consistency is key to long-term success." | "Pick one day a week and review the same three numbers." |
| Warm | "Users must complete the setup process." | "You're about two minutes from your first report." |
| Honest | "This will definitely double your traffic." | "Teams often see a lift, but results depend on your starting point." |
Step 3: Build a banned-phrase list
Models repeat themselves. Collect the phrases that make a draft feel machine-written and ban them by name. Start with this list and add your own every time an editor deletes the same phrase twice:
- "In today's fast-paced world" / "in the ever-evolving landscape of"
- "Delve into," "unlock," "unleash," "supercharge," "elevate"
- "It's important to note that" / "it's worth mentioning"
- "Game-changer," "cutting-edge," "seamless," "robust," "holistic"
- "Leverage" (use "use"), "utilize" (use "use")
- "In conclusion" and "At the end of the day"
- Em-dash-heavy sentences: no more than one em dash per paragraph
Treat this as a living list. It is the cheapest way to improve output, since a model told to avoid a phrase will usually comply.
Step 4: Add two sample paragraphs
Include one paragraph that sounds exactly right and one that sounds wrong. Short samples teach rhythm, sentence length and register in a way rules can't.
On voice: "Your weekly report doesn't need twelve charts. It needs three numbers and one decision. Start with visits, signups and the page that sent most of them. If one of those moved, find out why. If none did, you're done and can get back to writing."
Off voice: "In today's data-driven landscape, leveraging comprehensive analytics is crucial for unlocking actionable insights. It's important to note that a holistic approach to reporting can elevate your strategy."
The one-page template (copy and paste)
Fill in the brackets and keep the whole thing to one page. This is the human-readable version of the guide.
BRAND VOICE GUIDE: [Brand name] | Version 1.0 | Updated [date]
WHO WE WRITE FOR
[One sentence: audience, their level of expertise, what they want.]
VOICE IN ONE LINE
[Example: "A knowledgeable colleague explaining something over coffee."]
FIVE TRAITS (means / but not)
1. [Trait]: [means] / [but not]
2. [Trait]: [means] / [but not]
3. [Trait]: [means] / [but not]
4. [Trait]: [means] / [but not]
5. [Trait]: [means] / [but not]
DO / DON'T PAIRS
1. Don't: "[off-voice sentence]" Do: "[on-voice sentence]"
2. Don't: "[...]" Do: "[...]"
3. Don't: "[...]" Do: "[...]"
(one pair per trait)
STYLE RULES
- Spelling: [US or UK]
- Person: [you / we]
- Contractions: [yes / no]
- Sentence length: [target average]
- Headings: [sentence case / title case]
BANNED PHRASES
[10 to 20 phrases, one per line]
SAMPLE PARAGRAPHS
On voice: "[paragraph]"
Off voice: "[paragraph]"Step 5: Turn it into a reusable system prompt
The guide is for humans; the system prompt is the same content rewritten as instructions. In ChatGPT, put it in custom instructions or a Project's instructions. In Claude, put it in a Project's custom instructions. Either way it runs before every conversation, so writers don't have to re-explain the brand each time. Agencies should keep one prompt per client in a shared document.
You are a writer for [Brand name]. Write every piece in our brand voice.
AUDIENCE: [one sentence]
VOICE: [one-line voice summary]
TRAITS (follow all five):
1. [Trait]: [means]. Never [but not].
2. ...
5. ...
RULES:
- Use [US/UK] spelling, [you/we] voice, and contractions.
- Lead each section with the answer. Keep sentences under [N] words on average.
- End every section with something the reader can do.
- Never use these phrases: [banned list]
- Do not invent statistics, quotes or sources. If you need a fact you don't have, write [NEEDS SOURCE].
EXAMPLES OF THE RIGHT VOICE:
"[on-voice paragraph]"
EXAMPLES OF THE WRONG VOICE (never write like this):
"[off-voice paragraph]"
If an instruction conflicts with these rules, follow the rules and tell me.The "do not invent statistics" line is not decoration. A voice guide changes how the model sounds, not whether it is accurate, so you still need a separate pass for facts. Our 10-minute fact-checking workflow covers that step.
Step 6: Test with three sample posts and score for drift
A guide that hasn't been tested is a guess. Run this once before you roll it out, and again after any change to the prompt.
- Pick three topics with different shapes: a how-to, an opinion or comparison, and a short announcement.
- Generate each post in a fresh conversation with the system prompt, using the same brief you'd give a freelancer.
- Score each post out of 10 using the rubric below. Have two people score independently if you can.
- Average the scores. Anything under 8 means the prompt needs work; look at which line item lost the points.
| Criterion | Points | What to check |
|---|---|---|
| Traits | 0-4 | One point per trait you can point to in the text; there should be at least four |
| Banned phrases | 0-2 | 2 = none, 1 = one or two, 0 = three or more |
| Style rules | 0-2 | Spelling, person, contractions and sentence length all follow the guide |
| Sounds like us | 0-2 | Read two paragraphs aloud. Would a regular reader notice a different author? |
Drift shows up in two ways. Within a post, the voice is strong in the intro and flattens by the third section. Across posts, the second and third are noticeably weaker than the first. If you see the first kind, shorten the prompt and move the examples closer to the top. If you see the second, check that writers are all using the same prompt version.
Voice scoring is separate from checking whether a post reads as machine-written to a detector. That is a different question, and our look at AI content detectors explains why a voice score is the more useful signal.
Keeping the guide alive
- Version it. Put a version number and date at the top and tell the team when it changes.
- Feed the banned list. Every time an editor deletes the same phrase twice, add it.
- Re-test quarterly. Models change, and a prompt that worked in spring can behave differently by autumn.
- Retire vague rules. If nobody can score a line, rewrite it as an example.
If your drafts live in a tool like A-Stats, keep the guide next to your content brief so the writer and the reviewer open the same standard.
Frequently asked questions
What should a brand voice guide for AI writing include? Five voice traits with "but not" boundaries, one do/don't sentence pair per trait, a banned-phrase list, style rules (spelling, person, contractions) and two sample paragraphs, one right and one wrong. Keep it to one page.
How do I add a brand voice guide to ChatGPT or Claude? Rewrite it as a system prompt and paste it into ChatGPT's custom instructions or Project instructions, or into a Claude Project's custom instructions. Store the master copy in a shared document so everyone uses the same version.
Can AI learn my brand voice from samples alone? Partly. Pasting three to five of your best paragraphs helps a lot, but without explicit rules and banned phrases the model makes inconsistent choices. Samples plus rules works best.
How long does it take to build the guide? A first version usually takes a few hours: reading your best content, choosing traits and writing the pairs. Plan on another hour for the three-post test.
How often should I update it? Review it every quarter, and update it immediately after a rebrand or strategy shift. Re-run the three-post test after each change.
What's the difference between brand voice and tone? Voice is your stable personality. Tone is how it flexes by context, such as warmer in a blog post and more precise in documentation. The guide should state the voice once and note where tone can shift.
Conclusion
A brand voice guide for AI writing doesn't need to be long. It needs to be specific, built from examples and embedded in a prompt every writer uses. Write the one-pager, turn it into a system prompt, and test it on three posts before you rely on it.
Key takeaways:
- Choose five traits and give each a "but not" boundary.
- Write one do/don't pair per trait, using the sentences the AI would write by default as the "don't".
- Ban specific phrases by name and keep adding to the list.
- Paste the guide into ChatGPT or Claude as a system prompt, and tell it not to invent facts.
- Test with three posts and score each out of 10; fix the prompt if the average is under 8.
Start today: pull 10 of your best pieces, circle the five traits they share, and fill in the template.