Describing your voice produces generic writing with your adjectives applied to it.
"Friendly but professional, approachable, expert without being stuffy" describes almost every brand that has ever written a voice guideline. A model given that description writes the average of everything matching it, which is exactly the register you were trying to escape.
Showing works. Describing does not. And the most useful description is not of what you sound like — it is of what you never do.
Why descriptions fail
Voice adjectives are relative and the model has no reference point.
"Conversational" against what? Against a legal contract, a technical manual, and a friend's text message, "conversational" means three different registers.
Worse, the common adjectives are the ones every brand claims, so the model has seen them attached to an enormous range of writing and produces the middle of it.
Three descriptions that do carry information, because they are relative to something concrete:
- "Like our documentation, not like our ads"
- "Closer to a colleague explaining something than a company announcing something"
- "The register of someone who assumes you already know the basics"
Each names a comparison. That is what makes them usable.
What to supply instead
1. Three to five examples of your own writing
Real, published, and chosen because they are right rather than because they are recent.
Choose deliberately:
- One that shows how you open — usually the most distinctive thing about a voice
- One that shows how you explain something difficult
- One that shows how you disagree or say no, which is where voice is most visible
- One short piece, if you write short things
Say what to take from each. "Match how these open — the conclusion arrives before the explanation" is an instruction. "Match the tone" is not.
2. A never-do list
This is the highest-value item and almost nobody writes one.
A voice is defined more sharply by its exclusions than its qualities. Yours might include:
- No exclamation marks
- No rhetorical questions as openings
- Never "in today's fast-paced world", "unlock", "leverage", "seamless", "game-changing"
- No sentence beginning "whether you're a X or a Y"
- No urgency language on anything that is not urgent
- No claim we cannot support on the same page
- No em-dash-heavy sentence rhythm (if that is not yours)
Concrete, checkable, and it removes most of what marks generated copy — because the tells are largely a shared set of habits, and naming them excludes them at generation rather than at editing.
3. Three structural rules
Not tone. Structure — which is where voice actually lives.
Examples from this site's own writing:
- The answer comes first, in bold, before the explanation
- Sections stand alone. No "as mentioned above"
- Every claim states its source or is cut
Structural rules are followed more reliably than tonal ones, because they are unambiguous. A model can verify it did not write "as mentioned above". It cannot verify it sounded friendly.
The specification card
Everything above, on one page, pasted at the start of any writing task.
Voice specification
Examples of us getting it right: [three to five pieces, with a note on what to take from each]
Never: [the exclusion list]
Structure: [three rules]
Comparison: closer to [X] than to [Y]
Write it once. It takes an afternoon and it is reusable for everything. Where it fits in the wider approach.
Keep it under a page. A specification long enough to feel thorough stops being pasted.
Testing whether it worked
Three checks, in order.
1. The substitution test. Would this paragraph be equally at home on a competitor's site? If yes, the voice did not transfer — the register did.
2. The never-do check. Scan for everything on the exclusion list. This is mechanical and catches most failures.
3. Read it aloud. Voice is auditory and the eye is forgiving. A sentence that reads acceptably and sounds wrong is wrong.
If it fails, the fix is usually the examples rather than the instructions — a better-chosen example does more than another paragraph of description.
The limit worth knowing
A model can approximate how you write. It cannot supply what you would say.
Voice is not only register. It is what you are willing to claim, what you refuse to exaggerate, what you admit you do not know, and which things you think are worth an opinion. None of that is stylistic, and none of it comes from a specification card.
Which is the honest boundary: use the card for the scaffolding, and write the parts that carry a position yourself. A piece that sounds like you and says nothing you would say is worse than one that sounds generic, because it is a more convincing forgery of your judgement. Where the line sits.
A note on voice for email specifically
Email voice is more exposed than web voice, because it arrives in a personal space and is read faster.
Two things matter more there:
- The register of the opening line. A slightly-too-enthusiastic first sentence is the clearest tell there is
- Consistency with previous sends. A list learns your voice, and a drift in register reads as somebody else writing — or as a template
Include two previous emails in the specification if you write a lot of email. What email voice is doing.
Frequently asked questions
How do I get AI to write in my brand voice?
Show rather than describe. Supply three to five real examples with a note on what to take from each, a list of things you never do, and three structural rules. Voice adjectives like "friendly but professional" describe nearly every brand, so a model given them writes the average of all of them.
Why doesn't describing my tone of voice work?
Because tone adjectives are relative and the model has no reference point. "Conversational" means something different against a contract, a manual and a text message. Descriptions that name a comparison — "like our documentation, not our ads" — do carry information.
What is a never-do list?
The exclusions that define a voice: banned phrases, punctuation you avoid, sentence shapes you never use, claims you will not make. It is the highest-value item in a voice specification and almost nobody writes one, because a voice is defined more sharply by its exclusions than its qualities.
How do I know if the voice actually transferred?
Ask whether the paragraph would be equally at home on a competitor's site — if it would, you got the register rather than the voice. Then scan mechanically against the never-do list, then read it aloud, because voice is auditory and the eye is forgiving.
Should I include examples or just describe the style?
Examples, and say what to take from each. "Match how these open" is an instruction; "match the tone" is not. A single well-chosen example outperforms a paragraph of description and takes less time to supply.
Can AI fully replicate a brand voice?
It can approximate the register. It cannot supply what you would say — what you are willing to claim, what you refuse to exaggerate, what you admit not knowing. A piece that sounds like you and says nothing you would say is worse than a generic one, because it forges your judgement more convincingly.
The short version
- Choose three to five pieces of your own published writingChoose three to five pieces of your own published writing that are right rather than recent.
- Note what to take from eachNote what to take from each u2014 how you open, how you explain something hard, how you say no.
- Write a never-do listWrite a never-do list : banned phrases, punctuation, sentence shapes, claims you will not make.
- Write three structural rulesWrite three structural rules about how your content is built, not how it sounds.
- Add one comparisonAdd one comparison u2014 closer to X than to Y.
- Assemble it on one pageAssemble it on one page and keep it under a page.
- Paste it at the start of any writing taskPaste it at the start of any writing task
- Test the output with the substitution testTest the output with the substitution test u2014 would this be at home on a competitor's site?
- Scan mechanically against the never-do listScan mechanically against the never-do list
- Read it aloudRead it aloud , and if it failed, replace an example rather than adding description.
Free: The 60-Minute Email Authentication Fix
A no-fluff checklist to set up SPF, DKIM & DMARC correctly and pass Gmail & Yahoo's sender requirements.
Muhammad Basim has worked in digital marketing since 2013, focused on email deliverability and AI-assisted content production. He is the author of The Email Deliverability Playbook and The Email Copywriting Playbook.
Related Articles
Does Send Time Optimisation Work?
Send time optimisation models are trained on when recipients opened messages, and open timestamps now include machine activity that has nothing to do with when anyone was reading. Privacy protection pre-fetches message content on its own schedule, regardless of whether the recipient opened it. Security gateways fetch content during pre-delivery scanning, which happens before the […]
Prompting for Marketers: The Part That Actually Matters
Output quality is set mostly by the material you supply, not the phrasing you use. That is the finding people take longest to accept, because it is less interesting than the alternative. A carefully worded request with no context produces a competent generic answer. A plainly worded request with your actual customer language, your positioning […]
What You Grant When You Connect a Tool
Your security perimeter includes every vendor you have ever connected, including the ones you stopped using and never disconnected. A connection is a standing grant. It does not expire because you stopped logging in, it does not lapse because the trial ended, and it does not disappear when you delete the app from your phone. […]