Most brand voice guidelines fail the moment an AI tool touches them, and the failure has nothing to do with the model. It has to do with the format. Brightscout published guidelines this year built for both audiences at once, the team and the tools, and Contentoo's take lands on the same point from production: the context around the tool matters more than the model. We run a voice as a file every night, and the format decides everything. A model executes rules. It cannot absorb vibes.
What vibes look like versus what rules look like
Open a typical voice document and you find sentences written for humans who absorb tone by working nearby. "Sound conversational, not corporate." "Be warm but credible." A person on the team reads those, compares them against ten existing posts, and absorbs the pattern. A model reads them and produces the statistical average of its training data, which is where the generic corporate voice everyone complains about actually comes from.
The same voice, written as rules, looks different. Second person, present tense. Sentences under twenty words on average, with deliberate variance. No em dashes. No sentences starting with "So" or "And." Banned opener list. Three worked examples, before and after. Every line is executable: the model can follow it, check itself against it, and a reviewer can mark violations without arguing about taste.
The rule types that carry a voice
Four types survive contact with automation. Banned patterns come first: openers, filler phrases, punctuation tells, and constructions the brand avoids. Quantified constraints come second, because "short sentences" drifts and "under twenty words" does not. Worked examples come third, and negative examples carry more weight than positive ones, because a model copies structure faster than it infers taste. The fourth type is the decision log: when someone in the room rules a phrasing off-brand, the ruling joins the file, with the reason attached.
That last type is where voice actually lives over time. Typeface's orchestration approach enforces brand standards at generation time inside the tool, and enterprise teams need exactly that. A team without an enterprise platform can get the same effect from a text file that grows one ruling at a time, as long as every correction becomes a rule instead of a conversation.
A correction we shipped as a rule
Our own voice file changed this month for a reason that will sound familiar. A draft overused two sentence openers, and the critique came back with a specific list of the constructions to stop repeating. That critique became three lines in the file: the two banned opener patterns and a standing instruction to vary sentence openings. Every draft since has inherited the correction. Nobody had to remember it, mention it in a brief, or catch it again.
The scoring pass came with it. Before a draft ships, we check it against the file: the banned patterns, the length limits, the structural tells, one pass, a score, and a revision if the score lands low. The check runs the same way every night because the check is part of the file too.
Where the file lives decides whether tools follow it
A guideline in a PDF reaches nobody automatically. The same guideline in a repository, next to the drafts it governs, reaches every tool with file access, and the tools with file access are the ones writing the drafts now. Enterprise platforms solve the reach problem centrally: brand standards enforced inside the generation workflow. A plain text file next to the work solves it for the price of discipline, and it survives every platform switch, because the file travels while the tools rotate.
So audit your voice document the way you would audit code. Replace every adjective with a limit, every vibe with a pattern, and every meeting ruling with a line in the file. The voice that survives automation is the one a stranger could apply without ever meeting the team that wrote it.
Sources
The external viewpoints in this post come from these outlets:
- Brightscout published voice guidelines designed so a team and AI tools can follow the same document.
- Contentoo argued that the context teams build around AI tools matters more than the model choice.
- Typeface described enterprise enforcement of brand voice at generation time inside its platform.
The voice-file workflow, the banned-opener correction, and the scoring pass are ours, running nightly on this blog since November.

