Brand voice — the master prompt
Reading time: 20 minutes. Hands-on time: 30 minutes. What you'll have at the end: A brand voice document specific enough that AI-generated ads sound like you and not like a generic AI marketing template.
This is the most important tutorial in the library. Spend the time. Every single AI generation Scarif One produces — every ad, every email, every review reply, every social post — reads your brand voice document as the master prompt. Get it right and outputs feel inevitable. Get it generic and you'll spend the next year regenerating.
What "brand voice" actually means
In Scarif One, your brand voice is a markdown document, 250-500 words, that you can edit any time at /setup → Brand voice. It's not a colour scheme, not a logo, not a Pantone code. It's how you sound when you write.
When the AI gets a request like "generate a hero ad for our flagship coffee," it actually receives:
[brand voice document — 350 words]
+
[the product creative brief — 150 words]
+
[the user's specific request — 50 words]
That whole package becomes the prompt. The AI reads your brand voice first, the product context second, and the request last. The brand voice constrains everything downstream.
If your brand voice says "we never use the word 'premium'", the AI never uses "premium" in any ad it generates for you. That's the level of control you have.
The four paths to a brand voice doc
Pick the one that fits where you are, then come back here for editing.
Path 1 — AI scan (recommended for established brands)
What it does: Gemini 2.5 Pro fetches your live website, reads 5-15 pages (homepage, About, product pages, blog, taglines), and drafts a brand voice document grounded in your existing copy. Adds Google Search grounding for context.
When to use:
- ✅ You have a live website with substantial copy (>500 words across ~3 pages)
- ✅ Your existing copy is consistent enough to be a good source
- ❌ Your website is brand-new with placeholder text
- ❌ Your existing copy was AI-generated already (you'll get an echo of generic AI voice back)
How to run it:
/onboardingStage 3 → click AI scan card → paste website URL → Run- ~30 seconds
- ~£0.04 against your Gemini key
- Edit the result before saving
Path 2 — Paste a doc (for brands with an existing brand book)
What it does: a markdown editor. You paste your existing brand book / voice guide / copywriting guidelines and it gets saved as the master prompt.
When to use:
- ✅ You already have a 1-3 page brand voice document
- ✅ You've worked with a copywriter or brand strategist who delivered guidelines
- ❌ You have only a logo + colour palette (those aren't voice)
What format works: markdown (# headers, **bold**, - bullets) or plain text. Both fine. Keep it under 800 words — beyond that the AI starts to lose details.
Path 3 — 5-question interview (for brands without an existing voice)
What it does: AI asks 5 specific questions, you answer in plain English, AI drafts the doc.
The 5 questions:
- Describe your customer in one sentence.
- What does your customer say to themselves before buying?
- Three adjectives describing how you sound.
- Three adjectives you'd never want to sound like.
- Pick a competitor — what do they sound like that you don't want to?
When to use:
- ✅ Pre-launch / pre-website brand
- ✅ You know your customer and your stance but haven't written it down
- ❌ You're not sure who your customer is yet (do that first — see "Customer first" below)
This path produces the most consistent results when the answers are specific. "Coffee drinkers" → bad. "Specialty-coffee drinkers in their 30s who roast at home and won't drink supermarket coffee" → good.
Path 4 — Write your own
Empty editor. You know exactly what you want.
When to use:
- ✅ You're a copywriter or brand strategist who can write this in your sleep
- ✅ You've used Scarif One before and have a known-good doc to paste in
- ❌ Don't pick this path "to save time" if you're a first-time customer — you'll write something generic and the AI will reflect it back generically
What a great brand voice doc actually contains
After you generate (or write) the doc, edit it. Aim for these four sections:
Section 1 — Voice (3-5 sentences)
The high-level register. Two or three adjectives plus a sentence of context.
Generic, useless:
Friendly, professional, modern.
Specific, useful:
Warm but slightly defiant — a Brighton accent rather than mid-Atlantic. Anti-corporate. Founder-led, never marketing-team-led. Will use the word "knackered" without irony.
The test: read your version aloud. Does it sound like a real person describing a real brand? Or does it sound like a brief written by a committee?
Section 2 — Phrases we use (5-15 examples)
Specific phrases, not adjectives. Real strings the AI can latch onto.
Bad:
Use casual language.
Good:
- "Made for the morning person who refuses to compromise."
- "Brewed by humans, not robots."
- "Spend less time deciding, more time drinking."
- "Not a coffee snob — just a coffee believer."
- "Made on a Tuesday in our shed in Hove."
These become the AI's vocabulary. The first ad you generate after editing this section will feel transformed.
Section 3 — Phrases we never use (5-10 examples)
The negative space. Just as important as Section 2.
Examples:
- "Synergy", "leverage", "best-in-class" — anything that sounds like a McKinsey deck
- "Premium experience" — say what you actually mean
- "Crafted with care" — too generic, every brand says it
- "Award-winning" — earn the trust through specifics, not adjectives
- Anything with "elevate"
The AI takes these as hard constraints. It will work around them.
Section 4 — How we differ (2-3 paragraphs)
Pick a named competitor. Write 2-3 paragraphs explaining how you differ. Be unfair. This isn't a balanced market analysis — it's the AI's mental model of "who we're NOT."
Example:
We are not Costa Coffee. Costa is a brand built for people who don't really care about coffee — they care about the predictability of getting a cup that tastes the same in Stansted as it does in Sheffield. There's nothing wrong with that, but it's not us. We're for the person who'd rather discover a new bean from Ethiopia than drink a £4 mocha.
Costa would write "Christmas in a cup." We would write "the bean we roasted on Wednesday — try it before Friday."
This section often produces the most distinctive AI outputs. The contrast sharpens the voice.
Common mistakes (and the fix)
Mistake 1: Generic voice tokens
Your voiceTokens array (3-7 short tags shown in the dashboard) gets used in narrower prompt contexts. If they're generic, the AI's narrow contexts will be generic.
Bad: ["friendly", "professional", "modern"] — every brand on Earth could claim these.
Good: ["Brighton-coastal", "founder-led", "anti-corporate", "specific", "playful"] — reads as you, no one else.
Edit voiceTokens directly in /setup → Brand voice → Tokens. 7 max. Quality beats quantity.
Mistake 2: Too long
A brand voice doc over 800 words starts losing the AI's attention. The most-important constraints get drowned out by detail.
Fix: if your draft is 1,200 words, cut the bottom third. Keep the four sections above; drop anything that's repetition or filler.
Mistake 3: Conflicting signals
If your doc says "we're warm and approachable" in one paragraph and "we're sharp and uncompromising" in another, the AI averages them — and produces ads that are neither.
Fix: read your doc looking for adjectives that fight each other. Pick a side per dimension.
Mistake 4: Adjectives without examples
"We are warm" is a hint, not a constraint. The AI doesn't know what your warm sounds like vs. another brand's warm.
Fix: every adjective must be followed by a real-phrase example or a concrete don't.
Mistake 5: Forgetting to update after launch
Six months in, you've learned what works. Your brand voice doc from day 1 is now stale.
Fix: quarterly recurring task — open /setup → Brand voice → spend 15 minutes adding 3 new phrases that are working and 2 that aren't. Save. The AI improves overnight.
Customer first
If you're stuck writing a brand voice doc, the problem might be upstream: you don't have a clear customer yet.
Before you can write voice, you need:
- A specific person, not a demographic. "Specialty-coffee drinkers in their 30s who roast at home" beats "millennial coffee enthusiasts."
- What that person believes before buying. Their internal monologue — the thing they'd say to themselves at 6am with a half-asleep face.
- What they're sceptical of. Every customer comes pre-loaded with mistrust toward your category. Address it.
Spend 30 minutes thinking about your customer before you write voice. That 30 minutes pays back 100×.
The 5-question interview path (above) is built around this — the first two questions force you to articulate the customer.
Editing workflow
After you've saved a draft, dogfood it:
- Generate 3 ads with your current voice doc. Different concept types.
- Read them critically. Do they sound like you? Pick the worst-sounding one.
- Identify the line in your voice doc that's missing or wrong, given that bad ad.
- Edit the voice doc. Add the missing constraint or example.
- Regenerate the same ad. Compare.
- Repeat until the worst ad of three feels acceptable.
You'll iterate ~5-10 times in your first week. That's normal. The doc gets sharper each iteration.
After two weeks, you'll find yourself editing the doc once a month, then once a quarter. The voice stabilises as your AI memory accumulates.
When to hire a copywriter to do this for you
Honestly: at the seven-figure-revenue mark or when you have a marketing team who'll be using Scarif One, paying a brand strategist £1,500-£3,000 to write your voice doc is good ROI. They produce something tighter than the AI scan path can.
But for everyone else: write it yourself. The 30 minutes you spend writing the doc forces you to articulate things you'd otherwise let stay vague. That clarity has compound returns way beyond the doc itself.
Where to next
- Tutorial 04 — Your first ad — the deep dive (now that your voice is sharp, generate)
- Tutorial 06 — Drafting email campaigns (same voice doc, different output type)
- Tutorial 11 — The dashboard tour (see where the voice doc plugs into other features)
- Tutorial 12 — Your weekly strategy memo (the AI reads voice doc + analytics to recommend next moves)