Brand Voice Guidelines an AI Can Actually Follow

Brand voice guidelines work for AI when every line is a rule a machine can check: reading level, maximum sentence length, banned words, facts that must always appear, a hashtag ceiling per platform, an emoji policy, a claims policy, and one CTA pattern. Adjectives like bold and authentic do nothing. Replace each adjective with a number, a list, or a literal example.

There is a simple test. Hand your guidelines to a stranger and ask them to write two captions that both obey the document but read nothing alike. If they can do it, you have not written guidelines. A language model is that stranger every single time, with no memory of the feedback you gave last week.

Most brand books are written for a design review, not for a writer under deadline. They describe a feeling. A model has no feeling to match. It has a set of choices to narrow, and words like premium, human and authentic narrow nothing, because any caption you can imagine can be defended as authentic.

So the model falls back on its defaults, and its defaults are the average of the internet: rhetorical questions, three-word fragments for emphasis, a closing line that starts with "Because", and the word seamless. You get ten captions that could belong to any D2C brand in the country. The fix is not a cleverer prompt. The fix is a document that removes choices.

If two writers can follow your guidelines and produce opposite sentences, you have not written guidelines.

Every field below is a number, a list, or a literal sentence. None of them is an adjective. Write all eight and you have a document that survives contact with a generation tool.

Banned words need the near-misses too

A ban on "unlock" is easy to route around. The model writes "unleash", then "unlocks", then "the key to". Ban the family, not the word. Then ban the structures you dislike as well, because tone lives in structure: rhetorical questions as openers, one-word sentences for drama, and the phrase "we are excited to announce". A banned list that never grows is a list nobody is reading. Add an entry every time you edit the same thing out twice.

Always-say facts are your compliance layer

These are the sentences that must appear when the topic comes up, written exactly as you want them published. Price format. GST treatment. Shipping scope. Return window. Write them as sentences, not as notes. A note that says "mention GST" comes back as "plus taxes" or "GST applicable", which is not the same promise. A sentence comes back as the sentence.

The claims policy is the field that saves you

This is the one field with legal weight, and it is the field almost no brand guideline contains. State what the AI may assert without a source. For most brands that means: no numeric outcome claims, no superlatives such as fastest or best or number one, no competitor names, no health or income claims, and no statistics of any kind unless the number is written in this document. A festive discount claim must match the actual offer, in rupees, not a rounded-up version of it. If a claim needs a source, the source lives in the document or the claim does not get made.

One CTA, one shape

Pick a single pattern and enforce it: where the CTA sits, which verbs are allowed, and how many are permitted per post. Then list the CTAs you never want to see again. "Link in bio", "DM us for price" and "What do you think?" belong on most brands' banned list, and writing them down is faster than rejecting them one caption at a time.

What most brands have

We are bold, authentic and human. We speak like a friend, not a corporation. We are premium but approachable. We use humour where appropriate.

Four sentences and not one decision. "Humour where appropriate" is the worst line in it, because appropriateness is precisely the judgement you are asking the tool to make on your behalf. Give this to five writers and you get five brands.

The same brand, rewritten

Take a hypothetical Bengaluru brand selling filter coffee blends direct to customers. Same personality, written so a machine cannot misread it.

• Reading level: Class 8. Average sentence 12 to 14 words. Nothing over 22 words.
• Person: we for the brand, you for the reader. Never write about ourselves in the third person.
• Banned: artisanal, curated, elevate, indulge, journey, unlock, supercharge, coffee lover, perfect cup.
• Always say: "Roasted in Bengaluru." Once per product post, any position.
• Prices: always the rupee figure that appears on the site, with the symbol. Never "starting from" without the number.
• Hashtags: Instagram 8, LinkedIn 3, X 2. Lowercase except brand tags. Never inside a sentence.
• Emoji: none in captions. One in Stories text only.
• Claims: we may say the roast date is printed on every pack. We may not say a blend is the best or award-winning, and we make no health claim about caffeine.
• CTA: one, final line, imperative. Order, Try, Save, Reply. Never "link in bio".
• Humour: dry understatement about early mornings and traffic. No puns on brew or grind.

Every line there is falsifiable. You can hold a caption against it and say yes or no in two seconds. That is the entire bar a guideline has to clear.

1. 01Collect 20 posts you would happily publish again. Not the best performing ones. The ones that sound like you.
2. 02Measure them. Average sentence length, longest sentence, hashtag count, emoji count, where the CTA sits. Your document should describe what you already do well, not an aspiration you have never hit.
3. 03Highlight every word you would not say out loud to a customer standing in front of you. That is the first draft of your banned list.
4. 04Write five always-say facts as complete sentences. Cover price format, GST, shipping scope, and anything a customer or a regulator could hold you to.
5. 05Set one hashtag number and one caption-length target per platform. Instagram permits 2,200 characters and up to 30 hashtags. Permission is not a recommendation.
6. 06Write one CTA pattern and three banned CTAs.
7. 07Cap the whole thing at one page. If it does not fit, you are writing a strategy deck, not an instruction.

Give the same brief to your tool three times and keep all three outputs. Then hand the document and the captions to a colleague who did not write either, and ask them to mark each caption pass or fail, line by line against the fields. Two useful things fall out. You get captions you can ship, and you find every place where two reasonable people disagreed. Each disagreement is a missing rule. Add it, using the words you settled the argument with.

Run the test again whenever you add a platform. A voice tuned for a 3,000-character LinkedIn post breaks at 280 characters on X, and the break usually shows up as a dropped always-say fact.

The voice does not change by platform. The arithmetic does. Store the limits alongside the rules so the model is not guessing how much room it has.

A one-page doc in Google Docs works, as long as somebody pastes it into every prompt every time. That is exactly where it fails. Week three, a rushed festive post goes out without it, nothing bad happens, and the exception quietly becomes the norm.

The alternative is storing it once, in the place generation reads from automatically. In SocialPower that store is the Brand Brain: voice, audience, banned words, hashtag ceiling and claims policy, applied to every caption, hashtag set and CTA, on every connected account. Nobody has to remember to attach anything.

Two settings matter as much as the document itself. Autonomous Mode runs under an enforced daily post cap, and the approval gate is on by default, so a caption still passes a human before it publishes. If you are handing generation to any tool, read how the generation flow works before you switch it on, and check what each plan includes so you know the limits you are working inside. Paid plans start at ₹499 a month, excluding 18% GST.

Quarterly reviews of a voice document are theatre. Update it when the business changes: a new product line, a new platform, a price revision, a festive campaign that genuinely needs a different tone, or the third time you edit the same mistake out of a caption. That last trigger is the honest one. If you have deleted "seamless" three times, the fix is one line in the banned list, not a fourth deletion.

Keep a two-column changelog at the bottom of the page: date, and one line saying why the rule exists. In six months someone will ask why puns are banned, and the answer should already be in the file rather than in a manager's memory.