Guide / Brand voice

How to write a brand voice guide an AI can actually follow

A practical guide for Melbourne marketing and e-commerce teams. Use it whether or not you ever work with us.

Why most voice guides don’t work for AI

The typical brand voice guide was written for a human copywriter who already understands the brand. It lists three or four adjectives, maybe a “we are / we are not” table, and trusts the writer to fill the gaps with taste. AI tools don’t fill gaps with taste. They fill them with the most average version of what you asked for. A guide that works for AI is more specific, more concrete, and full of examples.

The seven sections that matter

1. Who you are talking to

Describe the reader as a person in a situation, not a demographic. “Someone buying a gift for a friend’s new apartment, unsure about their taste, wanting it to arrive before the weekend” gives a model far more to work with than “women 25–40, urban”.

2. Rules, not adjectives

Turn every adjective into rules you can check. “Warm” might mean: address the reader as “you”; one short personal sentence per email at most; no exclamation marks in service replies. “Confident” might mean: no hedging words like “perhaps” or “might” in product descriptions; state the material and the benefit, then stop.

3. Words you use and words you never use

Your own vocabulary: what you call your products, your customers, your sizes. And a banned list. Every AI tool has favourite filler words; list the ones you dislike and they will mostly disappear.

4. Structure by channel

For each channel, give the shape: an Instagram caption is a hook line, two sentences, a call to action, three hashtags maximum; a product description is a one-line summary, three benefit bullets, materials, care. Length limits in characters, not “keep it short”.

5. Ten to thirty approved examples

The single most useful section. Pick real pieces your team thinks are exactly right, a few per channel, and label them. Add two or three “not like this” examples with a note on what is wrong.

6. Facts and claims rules

Where product facts come from (the product data, not the writer’s memory), which claims need evidence, and which words you can’t use without it, such as “sustainable”, “hypoallergenic” or “Australian made”. The ACCC takes misleading environmental and origin claims seriously, and that applies to AI-written copy too.

7. Situations that change the tone

How the voice shifts for a delayed order, a complaint, a sold-out product, a sensitive date. Many brands sound wrong in exactly these moments because the guide only covered the happy path.

Testing your guide

Give the guide and one product to an AI tool and ask for a description. Then give the same guide to someone new in the team and ask for the same thing. Compare both with your best existing example. Wherever either drifts, the guide is missing a rule or an example. Two or three rounds of this usually produces a guide that works for people and machines alike.

Keeping it alive

Name an owner. Every time an editor rejects an AI draft for a voice reason, the reason goes into a running list. Once a month the owner turns repeated reasons into rules and swaps in fresh examples. A voice guide that hasn’t changed in a year is probably being ignored.

A one-page template

SectionFill in
ReaderOne paragraph describing the reader in a buying situation
Rules8–15 checkable rules
Our wordsProduct names, customer terms, size and colour conventions
Banned words15–30 words and phrases
Channel shapesStructure and character limits per channel
Examples10–30 approved, 2–3 rejected with notes
ClaimsSource of facts; words that need evidence
Tone shiftsDelays, complaints, sold-out, sensitive dates

If you would like a second pair of eyes on yours, or want it built into a content workflow, book a 15-minute call.