Brand Voice Belongs in a Check, Not a Prompt
An AI assistant told to write in a brand's voice will mostly comply. A fixed check — and a file that cannot be produced until the words pass it — is what makes it reliable.
The short answer: an instruction to "write in our brand voice" works most of the time. A fixed check gives the same answer every time it runs, and when it sits in the code that produces a file, an off-brand file cannot be made. If more than one company's name goes on what your assistant writes, you need the second one.
The example
In August, Cloudflare open-sourced Cloudflare OS, the AI workspace it built for its own staff. Our parent company, 2057 Holdings, deployed it for its operating companies — including us and our sister company, Safire Business Services.
We don't sound alike. Safire leads with the business outcome and ends with one clear next step. We lead with the answer, then the evidence, then how confident we are. Left alone, an AI assistant sounds like neither of us. It sounds like itself: fluent, upbeat, and generic.
What was built
Each brand in one place. Colors, fonts, logos, voice guide, and Office templates for each company, loaded before the assistant writes. This part is an instruction.
A fixed copy check. The assistant is instructed to run it before showing any draft. The check itself is not a second model's opinion — it is a set of rules kept in version control and applied identically every time:
- Brochure vocabulary: seamless, cutting-edge, best-in-class, game-changing, synergy.
- Fear-selling: "act now," "before it's too late."
- Overclaims: guarantees, "100% uptime," "zero risk."
- Client names, misspelled brand names, too many unexplained acronyms, and more than one exclamation point.
Each flag comes back with the excerpt and a specific fix.
Files that can't come out off-brand. Word documents and PowerPoint decks are built only from the company's official template — its own layouts, styles, logo, and footer. For files, the copy check is not an instruction: it runs in the code that builds them. If the words fail, there is no file.
No client material. Client mail, meetings, files, and projects are filtered out before the assistant receives them.
How to check your own assistant
- Where does your brand voice live? If the answer is "in the prompt," it's a preference.
- Is anything checked the same way every time? A second model reviewing the first is another opinion, not a check.
- Can an off-brand document still be produced? If yes, it eventually will be.
What we've confirmed
High confidence: for Word and PowerPoint files, the copy check and the template requirement run in code and are covered by automated tests that must pass before every deployment. A PowerPoint deck has been produced end to end, from the official template, through approval, to OneDrive.
Not yet exercised end to end: a Word document. We expect it to behave the same way; we haven't watched it yet.
Never assume, verify.