How to document brand voice so AI doesn't break your brand tone

Start with material, not adjectives

If you want AI to write “like the brand,” first collect real texts, not a list of traits like “professional” or “friendly.” According to TameTheBot, tone rules should come from evidence in samples, not from generalities, because without that the model starts guessing instead of imitating the brand's style.

The most practical starting point is 20-30 of your own communications: emails, support replies, offer descriptions, system error messages, and posts. In materials from Embark Agency and iContact the same direction repeats: first audit existing content, then name what actually works.

What must be in a brand voice guide

According to TameTheBot the document should be short but operational: include a summary of the voice, positions on a few tone axes, allowed and forbidden vocabulary, sentence mechanics, and a ready block of instructions to paste into a prompt. This approach aligns with iContact, where the voice chart describes not only a trait but also what it means in practice and what it does not mean.

In practice it's worth documenting:

• 3-5 brand voice traits, each with a brief description of behavior.
• A tone axis, for example formal vs. casual or serious vs. light.
• Say / don’t say: allowed words and constructions and forbidden ones.
• Syntax rules: sentence length, punctuation, use of the imperative, first vs. second person.
• 5-7 sample texts: a greeting, an error response, a sales email, a social post, a product description.

How to write rules so AI won't break them

Write rules as constraints rather than wishes. Instead of “sound natural,” better write: “use short sentences,” “don’t start with a definition,” “don’t use words X and Y,” “in technical replies give the solution first, then the explanation.” TameTheBot also emphasizes that each claim should be supported by a quote from samples; if something isn’t visible in the material, mark it as unconfirmed.

Good tests are straightforwardly practical. According to iContact and Moonb.io you should check whether the brand voice holds up in tougher situations: an error, a refusal, a support reply, a welcome email, or a pricing message. If the text still sounds recognizable after removing the logo, that’s a sign the voice guide works; if not, refine vocabulary, rhythm, and tone boundaries.

How to use the document with AI

The document should be “living” and available where AI actually writes, not just in a brand book on disk. Materials from Productivity and similar guides repeat the same rule: the guide should be saved where the model has constant access, not treated as a one-off presentation for the team.

A simple workflow works best:

• Paste a short summary of the voice guide into the prompt.
• Attach 2-3 exemplar fragments in the target style.
• Add a list of forbidden words and constructions.
• Force a revision if the text sounds too generic or too “marketing-y.”
• Update the guide every few weeks based on new examples.

Lub System helps B2B companies implement AI, automation and IT solutions end-to-end - from strategy to deployment. See our services or get in touch to discuss your case.