Content & Marketing
What Should Brand Voice Guidelines Say When Most Drafts Start With AI?
By Jim Vernon, Editor, AI Intelligence International · Published 27 August 2026 · Reviewed against our editorial standards · About the author
Most brand voice documents were written to influence humans who already had a voice of their own. They describe personality in adjectives — confident, friendly, human — which is close to useless as an instruction to a model and only marginally useful to a new writer.
This article covers how to rewrite voice guidelines as operational rules that survive being pasted into a system prompt, and how to keep them enforced once drafts start arriving pre-written.
Key takeaways
- Adjectives do not constrain output; rules, banned constructions and paired examples do.
- Every guideline should be checkable by someone who has never met your brand.
- Ship the guidelines as a reusable prompt block, not only as a PDF nobody opens.
- Enforce at review with a short checklist, because generated drafts drift toward the median every time.
Why do adjective-based guidelines fail?
Words like confident, approachable and expert have no consistent operational meaning. Ten writers produce ten interpretations, and a model produces the statistical average of how those words are used across the internet, which is precisely the generic register you were trying to avoid.
Worse, adjectives cannot be checked. A reviewer cannot say whether a paragraph is sufficiently confident, so the guideline never binds anything.
Replace each adjective with the behaviour it was standing in for. 'Confident' usually means 'state the recommendation before the caveats' — which is checkable.
What does an operational voice rule look like?
It names a construction to use or avoid and can be verified by reading. Examples: lead with the recommendation; never open with a rhetorical question; use second person for instructions; no more than one sentence of preamble before the substance.
Include a banned-phrase list. This is the single highest-leverage element, because the phrases that make writing sound generated are a fairly small, stable set, and removing them changes the register immediately.
Specify sentence-level defaults: preferred sentence length range, whether contractions are allowed, how numbers and currency are written, how you refer to the reader and to yourselves.
How do paired examples help?
A before-and-after pair conveys more than a page of description, for both humans and models. Show the generic version and your version of the same content, with a one-line note on what changed.
Six to ten pairs covering your common content types — a product announcement, a how-to opening, an error message, a sales email — will do more than any amount of prose.
Pairs also make review objective. A reviewer can point at the pair rather than arguing about taste.
How do you turn the guidelines into a prompt block?
Extract the rules and the banned list into a compact block that writers paste into every drafting session, alongside two or three of the paired examples most relevant to the format.
Keep it short enough to be used. A 300-word block that gets pasted every time beats a 2,000-word block that people skip because it slows them down.
Version it in the same place as the full document, and date it. When output quality shifts, the first question is which version of the block was in use.
What has to happen at review?
A fixed checklist applied to every piece: banned phrases absent, recommendation in the first paragraph, no unsupported statistics, examples specific rather than generic, and at least one detail that could only have come from your organisation.
That last item is the important one. Generated drafts are structurally competent and specifically empty, and specificity is the only reliable way to tell your content apart.
Timebox it. A voice review that takes twenty minutes per piece will be skipped under deadline; five minutes with a checklist will not.
How do you keep the guidelines alive?
Add to the banned list whenever a phrase starts recurring across drafts. It is a living artefact, and the drift is continuous because model outputs converge on common patterns.
Review the paired examples twice a year against recently published work, replacing any that no longer represent how you write.
Assign one owner. Voice documents maintained by committee become longer and less specific with every revision, which is the direction that destroys their usefulness.
Worked example: rewriting a voice document
A healthcare scheduling company had a nine-page voice document built on five adjectives. After adopting AI-assisted drafting, their content became noticeably uniform, and two customers commented that the blog sounded like everyone else's.
They rewrote it as a two-page operational document. Fourteen rules, a banned list of 31 phrases collected from three months of drafts, and eight paired examples.
The banned list included the constructions that had been appearing most: openings that began with 'In today's fast-paced', sentences of the form 'It's not just X, it's Y', the words leverage, robust, seamless and unlock, and any paragraph beginning with 'Whether you're'.
The rules included: state the practical recommendation in the first two sentences; every claim about scheduling outcomes must cite either their own anonymised data or a named source; use the word patient rather than user; no sentence over 30 words in an instructional paragraph.
They compiled a 280-word prompt block from it. Over the following quarter, review time per piece fell from roughly 40 minutes to 15, and time-on-page across new posts rose about 20%. The measurable change came almost entirely from the banned list and the first-two-sentences rule.
Frequently asked questions
How long should a voice document be?
Two to four pages for the full version and under 300 words for the prompt block. Longer documents are not read, and unread guidelines have no effect on output regardless of quality.
Should the banned list be shared with freelancers?
Yes, and it is usually the most valuable page you can give them. It communicates more about your register in one screen than the rest of the document combined.
Does a strong voice still matter if most readers arrive via AI answers?
It matters more for the readers who do arrive, since they are further down the funnel, and distinctive phrasing is one of the things that makes a source memorable enough to be sought directly.
Can the guidelines be enforced automatically?
The banned list and mechanical rules like sentence length can be checked by a simple script or linter, which catches most drift. Specificity and structure still need a human reviewer.