Brand voice guide: how to define a tone of voice an AI can actually learn

If you have ever pasted your brand voice document into ChatGPT and gotten back generic SaaS prose, this is for you. The problem is rarely the model. The problem is that 90% of brand voice documents were written to be read by a human copywriter who can interpret "playful but professional" — not by a model that needs concrete signal.

Four signals that actually move output

After training brand voice models for hundreds of teams, four inputs do almost all of the work. Skip the rest until these are tight.

1. 20–30 approved past posts

Not your best posts. Your most representative ones. Aim for a mix: 60% your standard cadence, 20% your sharper edge, 20% your warmest. Models learn shape from examples far better than from rules.

2. A "never say" list

Words and phrases your brand never uses. "Game-changing", "synergy", "leverage" (as a verb), "in today's fast-paced world". A 10-line ban list cleans up output more reliably than a 200-line style guide.

3. One concrete reader

Not a persona document. One sentence: who is the post for, what do they care about right now, what would make them roll their eyes. Specificity here is what makes the difference between "marketing copy" and "this person knows me".

4. Three forbidden structures

Most AI output reads generic because of structure, not vocabulary. Common offenders: the rhetorical question opener, the "X is more than just Y" frame, the three-word sentence for emphasis at the end. Name them, ban them, and 30% of the AI-tell disappears.

Format that travels

Lay it out as a short YAML- or JSON-shaped doc — even if a human writes it. The hierarchy makes it both human-skimmable and AI-parseable.

1. Voice attributes — 3–5 adjectives with one antonym each (e.g. "confident — not arrogant").
2. Reader — one sentence.
3. Example library — 20–30 posts with channel + post-type labels.
4. Never-say — bullet list, 8–12 entries.
5. Forbidden structures — 3 patterns with one example each.
6. House rules — formatting (capitalization, em-dash use, sentence-case headings).

Tune it like a model, not a manifesto

Brand voice documents are usually written, signed off, and frozen. That is a mistake when an AI is generating against them. Treat the voice file like a config: when a draft comes back wrong, find the missing signal and add it. After 6–8 weeks of weekly iteration, output stabilizes.

How Postify uses this

When you set up brand voice in Postify, the onboarding flow asks for exactly these four signals — past posts, reader sentence, never-say list, forbidden structures. The model retrains in 5–15 minutes. Every draft is then scored on voice fit; outliers route back for review before they reach the calendar.

The goal is not "AI that sounds like us". It is AI you barely have to edit — because the voice file gave it enough signal to be right the first time.