AI Brand Voice Guide: A One-Page Template for On-Brand Content

A few months ago the founder of a premium skincare brand we work with sent us two Instagram screenshots (details changed). One was from her account, the other from a competitor selling at half her price. Both opened with “Take your skincare routine to the next level!” Both were written by AI. Her message was one line: “Why do we sound as cheap as them?”

Her team had been handed a tool that writes 50 captions an hour, and no AI brand voice guide to keep it in check.

September 29 is Hari Sarjana Nasional, Indonesia’s National Graduates Day. If you’ve set aside budget for AI training this year, the most useful 90 minutes you can spend is building a one-page brand voice guide your team can paste into any AI tool. The template is below, and the workshop agenda is at the end.

Language models drift toward the average by default, and the average is the voice of every brand that has ever posted online. In almost every client content audit we run, we find the same pattern: instructions written as adjectives (“warm, professional, modern”) that the model fills with the most generic version it knows, no list of banned words, and everyone on the team prompting in their own style until the output sounds like three different brands. A one-page guide closes all three gaps because everyone pastes the same document at the top of every session.

---

Quick Summary

• A 30-page brand book gets read by nobody and cannot be pasted into a prompt. What you need is one page that fits inside a ChatGPT Project, Claude Project or Gemini Gem.
• The parts that do the most work: the “does not mean” line under each voice attribute, and a banned-words list where every ban has a replacement.
• Measure it with Voice Drift Rate: the share of AI assets that need a major rewrite. Our working threshold: above 25%, your guide is not doing its job yet.
• The expensive failure is an unchecked claim going live. In our cost model, Tier 3 review pays for itself if it stops one incident every five years; the formula is below so you can plug in your own numbers.

---

The One-Page AI Brand Voice Guide Template

In short: one document pasted at the top of every AI session, so everyone on the team writes inside the same guardrails.

Copy the block below, fill in the brackets, and keep it as one shared doc (Google Docs or Notion, either works). Keep it under 600 words so it fits in a ChatGPT Project or custom GPT’s instructions, a Claude Project, or a Gemini Gem. ChatGPT’s regular custom-instructions box is capped at 1,500 characters, so don’t use it for this. If your team has no access to Projects, paste the guide as the first message of every session.

BRAND VOICE GUIDE: [BRAND NAME]
Version [1.0] | Owner: [name, role] | Review: [every 3 months]

1. WHO WE ARE TALKING TO
Primary reader: [e.g. women 28-40, professionals in Jakarta and Surabaya,
have tried 3+ brands, skeptical of instant-result claims]
What they value: [e.g. clear ingredients, realistic results, their time]
What makes them leave: [e.g. hype, "bestie" slang, aggressive promos]

2. THREE VOICE ATTRIBUTES
Attribute 1: [Calm]
  Means: [short sentences, facts first, no exclamation marks]
  Does not mean: [cold or stiff like a bank brochure]
Attribute 2: [Honest about limits]
  Means: [we say who should NOT use the product]
  Does not mean: [doubting our own product]
Attribute 3: [Ingredient-literate]
  Means: [we state concentration, function, and how to use]
  Does not mean: [lab jargon with no explanation]

3. WORDS AND TONE
Form of address: ["you" in English; in Bahasa Indonesia always "Anda",
never "kamu", "kak" or "bestie"]
Words we use: [even, gradual, concentration, in-use testing, suitable for]
Banned word → replacement:
  guaranteed → in our in-use testing, [result + source]
  instant → visible from week [x]
  flawless / perfect / ideal → [name the specific result]
  next level → [name the concrete change]
  don't miss out → [promo end date, no pressure]
  miracle, revolutionary, glow-up, skin of your dreams → (cut, replace
  with an ingredient fact)
Punctuation: [no exclamation marks; no emoji in paid ads]
Tone by channel: [IG caption: lighter, 1 emoji allowed | Ads: no emoji |
Complaint replies: calm, apologize first, no promos]

4. CLAIM AND DATA RULES
- Every result number needs an internal source: write [SOURCE?] if missing.
- Banned: medical claims, "guaranteed", direct comparison with other brands.
- Never paste customer names, phone numbers, or chat logs into AI tools.
- Prices and promos come only from [link to official price sheet].

5. BASE PROMPT (paste at the start of every session)
[see the Base Prompt section below]

6. APPROVAL
Tier 1 (routine captions, comment replies): 1 editor. SLA 4 working hours.
Tier 2 (ads, product pages, bulk email, list prices from the price sheet,
  another party's name such as a collaborator or influencer):
  editor + brand lead. SLA 1 day.
Tier 3 (result/health/financial/comparison claims, new price promos,
  customer data, crisis): + legal/compliance. SLA 2 days.
No tier assigned = does not go live.

The “does not mean” line is the part teams skip, and the one that does the most work. Without it, AI reads “calm” as flat and “premium” as flowery. The negative line tells the model where to stop.

For a bilingual team, keep one guide with a word list per language (“level berikutnya” and “next level” both need listing).

The guide also helps with SEO: Google judges usefulness rather than the tool used, and a voice guide forces every asset to carry details only your brand has.¹ We cover the link to topical authority in our content marketing and blog SEO pillar.

---

Filling In the Three Attributes Without Guessing

Don’t invent attributes in a meeting room. Pull them from data you already have.

Collect your 20 best-performing posts from the last 12 months (judge by saves, DM replies, or conversions, not likes) and the 10 customer comments or reviews that most specifically praise how you communicate. Then score them against Nielsen Norman Group’s four tone dimensions: funny vs. serious, formal vs. casual, respectful vs. irreverent, enthusiastic vs. matter-of-fact.²

Dimension	Score 1–5	Example brand position	What it means for AI
Funny ↔ Serious	4	Leans serious	No jokes on product pages
Formal ↔ Casual	3	Middle	”Anda” in Bahasa, but short, spoken sentences
Respectful ↔ Irreverent	2	Respectful	No digs at competitors or customers
Enthusiastic ↔ Matter-of-fact	4	Matter-of-fact	No exclamation marks, no superlatives

Scale: 1 = left pole, 5 = right pole.

The three attributes usually fall out of this table.

One lesson from experience: if two senior people on your team score any dimension more than one point apart, stop. That gap means your positioning is not agreed yet, and no prompt will fix it. Settle it at the strategy level, then come back to the template.

---

The Word Lists: Fastest Payback in the Whole Template

The banned-words list has the best return of anything in the template. It takes an hour to build and the effect shows up in the very next output.

How to build it:

1. Ask the AI for 30 captions about your product with no guidance at all.
2. Highlight every phrase that makes you wince.
3. Any phrase that shows up three or more times goes on the banned list.
4. For each banned word, write the replacement you actually want. “Guaranteed” becomes “in our in-use testing, …” (for example, “in our in-use testing, 8 out of 10 users…”). “Instant” becomes “visible from week [x]”.

The replacements matter as much as the bans. A model given only a ban reaches for the nearest synonym: ban “flawless” and you get “perfect”. So every line on the banned list needs a replacement next to it.

---

The Base Prompt

Paste this at the start of every session, followed by the full guide and two samples of your brand’s best writing. It follows what OpenAI and Anthropic recommend in their prompting docs: clear instructions, context, and examples.³

You are a senior copywriter for [BRAND NAME]. The brand voice guide
below is mandatory. Follow it over any default style.

Hard rules:
1. Never use a word from the banned list. If tempted, use its replacement.
2. Never write a result number, percentage, or timeframe that is not in
   the brief. Write [SOURCE?] where that number would go.
3. Zero exclamation marks. Zero emoji unless the brief asks for them.
4. After writing, check your draft against the three voice attributes and
   report one line per attribute: pass / fail + reason.

On-brand writing samples:
[paste 2 best samples]

Task: [asset type] for [channel], goal [desired action],
length [x words]. Write 3 variations.

Rule 4 makes the model grade its own work. That grade does not replace your editor, but it usually saves you the first revision round, because the crudest mistakes get caught before a human sees the draft.

A technical note: in long conversations the model gradually loses hold of its opening instructions. In our client work, for a batch of 20 captions, starting a new session every 8–10 assets keeps output close to the guide.

---

Tiered Approval: Who Signs Off on What

For every asset, ask one thing: if this is wrong, what does it cost? Approval tiers map that answer to a review cost.

Tier	Example assets	Reviewer	Trigger to move up a tier
1	Organic captions, comment replies	Editor	A list price or another party’s name: move to Tier 2. Any claim: straight to Tier 3
2	Meta/TikTok ads, product pages, bulk email, list prices from the price sheet, a collaborator’s or influencer’s name	Editor + brand lead	A result, health, financial, or comparison claim, or a new price promo: move to Tier 3
3	Result/health/financial/comparison claims, new price promos, crisis comms, anything with customer data	+ Legal/compliance	Highest tier

Two risks fast-moving teams tend to miss.

Personal data. Pasting a customer list, WhatsApp chat exports, or phone numbers into a public AI tool is processing personal data under Indonesia’s Personal Data Protection Law (UU No. 27 of 2022).⁴ Because most public AI tools process data on servers outside Indonesia, it also touches the Law’s cross-border transfer rules (Article 56). Block 4 of the template exists to close that gap. If your team genuinely needs AI to process customer data, use a business or enterprise account with a data processing agreement, and bring in your DPO or legal counsel.

AI labels. Meta applies an “AI info” label to images it detects as, or that are disclosed as, AI-generated.⁵ For a premium brand, a labeled AI product shot can read as “they don’t have real photos”. Decide the policy once, write it into the guide, and your team stops guessing.

---

Before and After: A Premium Niacinamide Serum

Illustrative brand: “Arunika”, a premium local skincare label in Jakarta, selling a 30 ml serum at Rp389,000. A three-person content team produces around 80 assets a month with AI.

Before the guide, the team’s prompt was: “Write an Instagram caption for our niacinamide serum, premium and engaging tone.”

Take your skincare routine to the next level! Our newest Niacinamide
Serum is the perfect solution for the flawless, glowing skin of your
dreams. Formulated with carefully selected premium ingredients, it's
guaranteed to fade dark spots in just 7 days. Don't miss out!

Editor’s diagnosis: six banned phrases (next level, perfect, flawless, skin of your dreams, guaranteed, don’t miss out), not one fact about the ingredient, and “guaranteed to fade dark spots in just 7 days” is a result claim with no evidence behind it. In a paid ad, a claim like that risks platform rejection and runs into BPOM Regulation No. 3 of 2022 on the Technical Requirements for Cosmetic Claims, which does not allow cosmetic claims that imply a medicinal effect.⁶ Verdict: full rewrite.

After the guide, using the base prompt and two Arunika writing samples. Brief supplied: 5% niacinamide, 30 ml, Rp389,000, 2-3 drops at night after toner, not for irritated skin or right after a peel. Internal in-use note: more even tone reported from week 4 (report not final).

Caption (ready to post once [SOURCE?] is filled):

The 5% niacinamide in this serum works slowly, and that is on purpose.
In our internal in-use testing [SOURCE?], users started reporting a
more even skin tone from week 4 of daily use.

Use at night after toner, two to three drops.
Skin irritated, or just had a peel? Hold off until it settles.

30 ml, Rp389,000. Available on our site and official store, link in bio.

Model self-check (not for posting):

Self-check: Calm (pass, no exclamation marks) | Honest about limits
(pass, says when not to use it) | Ingredient-literate (pass,
concentration + how to use).

Look at the [SOURCE?] marker. “Week 4” was in the brief, so the model used it. But the brief also said the report is not final, and block 4 requires a source for every result number, so the model tagged it instead of presenting it as settled fact. This result claim automatically moves to Tier 3: the brand lead attaches the test data and legal/compliance signs off on the claim wording before it goes live. If the report never gets finalized, cut the sentence. The caption then drops back to Tier 2, since only the list price is left. For the principles that make premium buyers trust copy like this, see premium pricing psychology and psychology-based copywriting.

---

The Math: Voice Drift Rate and the Cost of Risk

Measure the guide with one formula:

Voice Drift Rate (VDR) = AI assets needing a major rewrite ÷ total AI assets × 100%

Major rewrite = editor changes more than 30% of sentences, OR a banned
word appears, OR a claim has no source and is not flagged [SOURCE?].

The numbers below are an illustration for Arunika. Your inputs will differ, so swap in your own.

Item	Before the guide	After 6 weeks
AI assets per month	80	80
VDR	55%	12%
Average edit time per asset	40 min	15 min
Senior editor hours on edits	53.3 hrs	20 hrs
Editor cost (Rp18M salary ÷ 160 hrs = Rp112,500/hr)	Rp6,000,000	Rp2,250,000

That is roughly Rp3.75 million saved a month, or 33 senior-editor hours back for strategic work. Worth having, but small next to what one bad claim costs. Here is the expected loss from claims that slip through:

Expected monthly loss = assets per month × probability a bad claim gets through × cost per incident

Without tiered approval   : 80 × (1/1,000) × Rp150,000,000 = Rp12,000,000
With guide + Tier 2/3     : 80 × (1/5,000) × Rp150,000,000 = Rp2,400,000
Difference                :                                  Rp9,600,000

Tier 2 review cost per month = 30 assets × 0.25 brand-lead hrs × Rp150,000 = Rp1,125,000 (assumption)
Tier 3 review cost per month = 10 assets × 1 legal hour × Rp250,000        = Rp2,500,000
Total review cost                                                           = Rp3,625,000

Minimum per-asset risk reduction for Tier 3 to break even
  = Tier 3 review cost ÷ (assets per month × impact)
  = Rp2,500,000 ÷ (80 × Rp150,000,000) ≈ 1 in 4,800

That last line matters most, because it doesn’t depend on the made-up probabilities, only on the Rp150 million impact assumption. Our assumed reduction is 1 in 1,250 per asset, nearly four times that threshold. Put in time terms: Tier 3 pays for itself if it prevents one incident roughly every five years. Add the Tier 2 cost and the threshold is about 1 in 3,300, or one incident every three and a half years.

The Rp150 million incident cost is a blended assumption: an ad campaign halted mid-flight, reprinted materials, and team hours spent on damage control. The 1/1,000 and 1/5,000 probabilities are illustrative assumptions too. The Rp250,000 hourly rate is for in-house legal (roughly a Rp40 million monthly salary); external counsel can bill several times that, and the break-even logic still holds.

To get your own numbers, count the assets that were pulled, rejected by a platform, or corrected by legal in the last 12 months, and divide by total assets in that period. To connect numbers like these to CAC and payback, see our guide to CAC, LTV and ROAS.

---

When This Template Won’t Work

It won’t rescue a brand whose positioning isn’t agreed. If the founder and the head of marketing picture different customers, AI will simply scale that confusion, fast.

It fails if you treat it as a banned-words list and nothing more. Teams that only strip banned words produce content that is clean and bland, because the character comes from the attributes and the writing samples.

For highly sensitive Tier 3 content, such as a crisis statement or a price increase announcement, we still recommend a human writes the first draft. Bring AI in afterwards, as an editor.

And an unmaintained guide goes stale. Schedule a review every three months, on the same rhythm as the content calendar in our content consistency guide.

---

National Graduates Day: A 90-Minute Team Workshop

Here is the agenda we have run with client marketing teams.

Pre-work (3 days before): one person pulls the 20 best-performing posts, 10 customer reviews, and 30 unguided AI captions.

Minutes	Activity	Output
0–15	Review the 30 unguided AI captions from pre-work together	First-read notes on the “before” set
15–40	Score the 4 tone dimensions, agree on 3 attributes	Block 2 of the template
40–60	Highlight clichés from those captions, build word lists	Block 3 of the template
60–80	Rewrite with the base prompt, compare results	Internal before/after example
80–90	Agree approval tiers and the document owner	Block 6 + a named owner

Save the result in a shared folder and book the first review six weeks out, when your first VDR number is in.

---

How Eranya Digital Helps

Some clients build this template themselves and stop there. Others come to us once they are shipping hundreds of assets a month across several channels and need a system: the voice guide, a prompt library per asset type, a logged approval flow, and a monthly VDR number management can read.

If you want an outside read on your last 20 AI-written assets, with a VDR score and a list of the claim risks we find, book a strategy session with our team.

---

References

---

<script type="application/ld+json">
[
  {
    "@context": "https://schema.org",
    "@type": "Article",
    "headline": "AI Brand Voice Guide: A One-Page Template for On-Brand Content",
    "description": "A one-page AI brand voice guide template: voice attributes, do/don't word lists, base prompts, approval tiers, and a before/after example to keep AI on-brand.",
    "image": "https://eranya.digital/images/blog/kesalahan-konten-lokal-website.webp",
    "author": {"@type": "Organization", "name": "Eranya Digital", "url": "https://eranya.digital"},
    "publisher": {"@type": "Organization", "name": "PT Eranya Digital Nusantara", "logo": {"@type": "ImageObject", "url": "https://eranya.digital/images/logo.png"}},
    "datePublished": "2026-09-29",
    "dateModified": "2026-09-29",
    "mainEntityOfPage": {"@type": "WebPage", "@id": "https://eranya.digital/blog/ai-brand-voice-guide/"},
    "inLanguage": "en-US",
    "keywords": ["ai brand voice guide", "brand voice guide template", "brand voice prompt for chatgpt", "keep ai content on brand", "tone of voice guidelines", "ai content approval workflow"],
    "articleSection": "Strategi Digital"
  },
  {
    "@context": "https://schema.org",
    "@type": "FAQPage",
    "mainEntity": [
      {"@type": "Question", "name": "What is the difference between brand voice and tone of voice?", "acceptedAnswer": {"@type": "Answer", "text": "Brand voice is the fixed personality: how your brand thinks and speaks on every channel. Tone is how that voice adjusts to the situation. The same brand can be lighter in an Instagram caption and calmer when answering a complaint, but word choice, attitude to claims, and how it addresses customers stay the same. In a one-page guide, write voice as fixed attributes and tone as per-channel rules."}},
      {"@type": "Question", "name": "How do I get ChatGPT or Claude to write in my brand voice?", "acceptedAnswer": {"@type": "Answer", "text": "Adjectives like warm and professional barely move the output. Constraints do: three voice attributes, each with a means and a does not mean line, a list of required and banned words with replacements, claim rules, and two or three samples of your own best writing. Paste all of it into your Project, custom GPT, or Gem instructions (or the first message of the chat), and start a fresh session every 8-10 assets so the model does not drift."}},
      {"@type": "Question", "name": "Is AI-generated content bad for SEO?", "acceptedAnswer": {"@type": "Answer", "text": "Not automatically. Google says it rewards helpful, high-quality content however it is produced. What breaks its spam policies is mass-produced content made mainly to manipulate rankings. Where AI content does fall short is when it carries none of your own experience, data, or point of view: it looks like thousands of similar pages and rarely builds authority."}},
      {"@type": "Question", "name": "Who should approve AI-written content before it goes live?", "acceptedAnswer": {"@type": "Answer", "text": "It depends on the risk. Routine captions need one editor. Paid ads, product pages, emails to your full database, and anything quoting an official list price need an editor plus the brand owner. Anything with result, health, or financial claims, new price promos, competitor comparisons, or customer data needs an extra legal or compliance review. Write these tiers into the brand voice guide so nobody has to guess."}}
    ]
  }
]
</script>

Footnotes

1. Google Search Central. Google Search’s guidance about AI-generated content. developers.google.com/search/blog/2023/02/google-search-and-ai-content. ↩

2. Nielsen Norman Group. The Four Dimensions of Tone of Voice. nngroup.com/articles/tone-of-voice-dimensions (the four-dimension framework: funny/serious, formal/casual, respectful/irreverent, enthusiastic/matter-of-fact). ↩

3. OpenAI. Prompt engineering guide. platform.openai.com/docs/guides/prompt-engineering (clear instructions, context, and examples in prompts); Anthropic. Prompt engineering overview. docs.anthropic.com/en/docs/build-with-claude/prompt-engineering/overview. ↩

4. Republic of Indonesia. Law No. 27 of 2022 on Personal Data Protection (UU PDP). peraturan.bpk.go.id/Details/229798/uu-no-27-tahun-2022. ↩

5. Meta. Our Approach to Labeling AI-Generated Content and Manipulated Media. about.fb.com/news/2024/04/metas-approach-to-labeling-ai-generated-content-and-manipulated-media (updated July 2024: the “Made with AI” label was renamed “AI info”). ↩

6. Indonesian Food and Drug Authority (BPOM). Peraturan BPOM Nomor 3 Tahun 2022 tentang Persyaratan Teknis Klaim Kosmetika (Technical Requirements for Cosmetic Claims). peraturan.bpk.go.id/Home/Details/223958/peraturan-bpom-no-3-tahun-2022. ↩