TL;DR: "Write in our brand voice" fails because a model has no idea what that means, and adjectives like friendly or bold describe every brand equally. The prompts that actually work extract a voice from real examples and turn it into testable rules, then apply and audit against those rules. Below are 25 of them, organized by job.
Why does "write in our brand voice" never work?
Type that instruction into any model and it produces something reasonable-sounding and completely generic, because the model was never given anything to check its own output against. "Friendly" could mean five different things depending on which brand said it. "Professional" describes a law firm and a fintech app equally well. An adjective is an impression, not an instruction, and a model can only follow instructions.
Most brand voice documents make this worse rather than better, because they were written for a person, not a model. A style guide that says "warm but authoritative" is genuinely useful to a new hire who can read between the lines using years of context about what the brand actually sounds like. A model has none of that context, so it fills the gap with whatever "warm but authoritative" statistically means across every brand that has ever used those two words, which is to say the industry average. Two competitors with opposite guides and the same adjectives will get output from the same model that sounds almost identical.
The fix is not a longer adjective list. It's a different kind of rule entirely: one you can point at a sentence and answer yes or no. "Contractions always" is testable: you can check any sentence and know immediately whether it complies. "Sounds warm" is not; two people will disagree about the same sentence. The prompts below are built to produce the first kind, never the second.
What makes a brand-voice rule testable, and why does it matter?
A testable rule names a concrete, checkable behavior: a word count ceiling, a banned word list, a rule about how many exclamation points are allowed on a page, a rule about what a sentence opens with. An adjective names a feeling and leaves the behavior to guesswork. The practical difference shows up the moment you try to use either one for anything besides generating a first draft.
| Feature | Adjective instruction (friendly, bold, professional) | Testable rule (extracted from your own examples) |
|---|---|---|
| A new writer can apply it without training | No, depends on that writer's private sense of bold | Yes, the rule states the exact behavior |
| The model can check its own output against it | No, nothing to check | Yes, pass or fail per sentence |
| Survives a switch between models or tools | Partially, each model's default friendly differs | Yes, the rule is external to any model's defaults |
| Catches drift six months later | No, impressions are not comparable over time | Yes, the same rule applies unchanged |
| Where it comes from | Guessed or copied from a style guide | Extracted from writing that already worked |
That's the mechanic behind every prompt in this post: extract from real writing, state the result as a rule, apply the rule, then audit against the rule. Not one of the 25 asks a model to be anything. Each one asks it to do something specific and checkable, organized below by the job it does: extraction, encoding, applying across formats, auditing, edge cases, and multi-brand work.
How do you extract a brand voice from writing you already have?
Extraction starts with three to five pieces of copy that already sound like you: an email, a landing page section, a support reply, whatever you'd point to and say "yes, that's us." Paste them into a model and ask it to describe patterns, not vibes. Favor pieces from different formats over five variations on the same email, since a rule extracted from one format alone tends to be a formatting habit rather than a voice trait, and pick recent copy over the brand's earliest writing, since voice tends to have moved on since year one.
1. Sentence rhythm and length
Below are [3-5] samples of writing from [BRAND NAME] that we consider strongly on-voice.
[PASTE SAMPLES HERE]
Analyze sentence structure only. Report:
- Average sentence length in words, and the range
- Whether sentences are mostly simple, compound, or a mix
- How often (if ever) a sentence starts with "And," "But," or a conjunction
- Whether paragraphs are typically 1-2 sentences or longer
Do not comment on tone or personality. Report only measurable structural patterns.
2. Vocabulary fingerprint
Same samples as above: [PASTE SAMPLES HERE]
List:
- 10 words or short phrases that appear more than once and feel distinctly "us"
- Whether we use contractions (yes/no/mixed, with examples)
- Whether we address the reader as "you" or in third person
- Any industry jargon we use versus jargon we visibly avoid
Output as two columns: "Words/patterns we use" and "Words/patterns we don't use," based only on evidence in these samples.
3. What the brand never says (the negative space)
Here is our approved writing: [PASTE SAMPLES HERE]
And here is writing from a competitor or an earlier draft that felt off-brand: [PASTE COMPARISON TEXT HERE]
Compare the two sets and list specifically what appears in the second set but never in the first: banned words, sentence patterns, structural habits, openers, closers. I want the negative space — what we don't do — not a general comparison.
4. How the brand handles bad news
Here are [2-3] examples of how [BRAND NAME] has communicated a delay, an outage, a price increase, or a mistake: [PASTE EXAMPLES HERE]
Extract the pattern as a rule, not a summary. Specifically:
- Does the message lead with the bad news or with context first?
- Is there an apology, and if so, how many words does it take?
- What happens in the sentence immediately after the bad news — a fix, a timeline, an apology, or something else?
- Any words this brand visibly avoids in bad-news writing (e.g. "unfortunately," "we apologize for any inconvenience")
5. Opening and closing patterns
Samples: [PASTE 3-5 SAMPLES HERE]
Look only at the first sentence and the last sentence of each piece. Report:
- Common patterns in how pieces open (a question, a claim, a scene, a stat)
- Common patterns in how pieces close (a call to action, a summary, an open question)
- Any opener or closer type that never appears, based on this evidence
6. Extracting from a single flagship piece
Sometimes you have one piece of writing everyone agrees is exactly right, and nothing else to compare it to yet. This prompt gets a usable first pass out of a single sample, with the honesty about generalization built in.
Here is one piece of writing from [BRAND NAME] that we consider our best example of the voice we want: [PASTE SINGLE SAMPLE HERE]
Extract as many specific, checkable patterns as this one sample actually supports — sentence length, vocabulary, structure, point of view. Flag clearly which patterns are strongly supported by repetition within this piece versus which are a single occurrence and should be treated as a hypothesis to confirm against a second sample, not a confirmed rule yet.
How do you turn extracted patterns into a reusable voice block?
Extraction gives you observations. The next four prompts turn those observations into a block you can paste at the top of any prompt, in any tool, for anyone on the team. That's the difference between a one-off analysis and something that survives being reused a hundred times: an analysis lives in the chat where you ran it, while a block is a standalone piece of text built to travel, get pasted, and get edited by someone who wasn't in the room when it was written.
7. Convert patterns into testable rules
Here are the patterns extracted from our writing: [PASTE EXTRACTION OUTPUT FROM STEP 1-6 HERE]
Rewrite each pattern as a single testable rule — a statement specific enough that you or I could check any given sentence against it and get the same yes/no answer. Bad example: "sounds confident." Good example: "never opens a sentence with 'I think' or 'we believe.'" Number the rules and group them under: Structure, Vocabulary, Point of view, What we never do.
8. Build the parameterized voice block
Using these rules: [PASTE RULES FROM STEP 7 HERE]
Assemble a single reusable block I can paste at the top of any generation prompt, in this format:
BRAND VOICE — [BRAND NAME]
Structure rules: [list]
Vocabulary — always: [list]
Vocabulary — never: [list]
Point of view: [rule]
Exceptions: [leave blank for now — I'll fill this in separately]
Keep every line as a rule, not a description. If a pattern from my rules doesn't fit one of these categories cleanly, put it under Structure rather than dropping it.
9. Compress into a short-context version
For a tool or a teammate who won't paste a full block every time, this trims the voice down to what actually changes model behavior.
Here is our full brand voice block: [PASTE FULL BLOCK FROM STEP 8 HERE]
Compress this to the 5 rules that would change the model's output the most if removed, ranked by impact. This shorter version needs to fit in a single short paragraph and still be enough to catch the most common off-voice mistakes. Drop anything that's true but wouldn't actually be violated in practice.
10. Find contradictions in the rule set
A voice block written in one sitting can quietly contradict itself: "always concise" next to "always explain the reasoning" is a fight waiting to happen the first time both apply to the same sentence.
Here is our brand voice rule block: [PASTE FULL BLOCK HERE]
Check this rule set for internal contradictions — cases where following one rule would require breaking another in the same sentence. For each contradiction found, state both rules involved and suggest which one should take precedence, with a one-line reason.
How do you apply the same voice block across email, landing pages, docs, social, and support?
The voice block doesn't change between formats. What changes is the register: how formal, how long, how much room there is for a joke. That's a dial you turn on top of a fixed voice, not a reason to rewrite the rules each time. This is the actual distinction between brand voice and tone of voice: voice is who you consistently are, tone is how loud or quiet you are about it in a given moment, and confusing the two is why some teams end up with a different-sounding voice document for every channel instead of one voice with five settings.
| Format | What stays fixed (voice) | What typically flexes (register) |
|---|---|---|
| Vocabulary, sentence rhythm, point of view | Subject-line length, greeting formality | |
| Landing page | Banned words, structural habits | Sentence length (often shorter than email) |
| Docs / help center | Point of view, vocabulary | Warmth (docs usually run drier) |
| Social post | Vocabulary, what the brand never says | Length cap, hashtag/emoji use |
| Support reply | Point of view, bad-news handling rule | Empathy phrasing, response length |
Each of the five prompts below takes the same voice block as input and names explicitly what's allowed to flex for that format, so the model isn't guessing at the register on top of a voice it's already following correctly.
11. Email
Brand voice block: [PASTE VOICE BLOCK HERE]
Write a [length, e.g. 150-word] email about [TOPIC]. Audience: [WHO]. Goal: [what the reader should do after reading].
Apply every rule in the voice block. Where a choice isn't covered by the block, default to shorter over longer and specific over general.
12. Landing page section
Brand voice block: [PASTE VOICE BLOCK HERE]
Write a landing page section for [FEATURE OR OFFER]. Structure: one headline, one supporting sentence, three benefit bullets, one call to action.
Apply the voice block's vocabulary and structure rules strictly. Landing page copy runs shorter than our email copy — keep sentences under [N] words even where the voice block's general rule allows more.
13. Documentation / help center article
Brand voice block: [PASTE VOICE BLOCK HERE]
Write a help center article explaining how to [TASK]. Keep the vocabulary and point-of-view rules from the block. Docs register: drop any rule about warmth or personality flourishes from the block and prioritize being scannable — numbered steps, one action per step, no rhetorical questions.
14. Social post
Brand voice block: [PASTE VOICE BLOCK HERE]
Write a [PLATFORM] post about [TOPIC/ANNOUNCEMENT], under [N] characters. Keep the block's vocabulary rules and "never say" list exactly. State explicitly whether the block's structure rules (sentence length, opener pattern) apply at this length, or whether they need to compress — and show your compressed version, not just the original rule.
15. Support reply
Brand voice block: [PASTE VOICE BLOCK HERE]
Customer message: [PASTE CUSTOMER MESSAGE HERE]
Write a reply using the block's vocabulary and point-of-view rules. This is a support context: prioritize the block's bad-news-handling rule over any rule about being upbeat or promotional, and do not include a sales mention unless the customer's own message asks about pricing or upgrading.
How do you audit whether a draft actually matches the voice?
A "does this sound like us?" read is subjective and inconsistent between two people on the same team. An audit against stated rules gives the same answer regardless of who runs it, which is the entire point of writing the rules down in the first place. This is also where the testable-rule mechanic pays for itself: an audit against adjectives produces an opinion, and two reviewers can disagree about the same draft and both be right by their own reading. An audit against a rule produces a verdict, checkable by anyone who reads the same rule the same way.
16. Line-by-line rule audit
Brand voice block: [PASTE VOICE BLOCK HERE]
Draft to check: [PASTE DRAFT HERE]
Go through the draft sentence by sentence. For each rule in the voice block, report: Pass, Fail, or Not applicable to this draft. For every Fail, quote the exact sentence that violates it and name which rule it breaks. Do not soften a fail into "mostly fine" — report it as a fail.
17. Blind distinguishing test
Here are two drafts of the same piece: [PASTE DRAFT A] and [PASTE DRAFT B]
Without telling you which is which, I'll tell you: one was written to match our brand voice block, one was not. Read both and tell me which one you believe is the on-voice draft, and list the three specific pieces of evidence (word choices, sentence patterns, structural choices) that led to your answer.
18. AI-tell scrub specific to this voice
Brand voice block: [PASTE VOICE BLOCK HERE]
Draft: [PASTE DRAFT HERE]
Separately from voice compliance, flag any generic AI writing patterns in this draft: symmetrical "not just X, but Y" constructions, a summary paragraph that restates what was just said, hedging phrases our voice block doesn't call for, or an ending question the block doesn't ask for. List each instance with its sentence.
19. Score before publishing
Brand voice block: [PASTE VOICE BLOCK HERE]
Draft: [PASTE DRAFT HERE]
Score this draft against the voice block: for each rule category (Structure, Vocabulary, Point of view, Exceptions), give a pass count out of the rules that applied to this draft, and one sentence naming the single biggest violation, if any. End with a clear ship or revise recommendation — no "it depends."
What happens to brand voice in apologies, refusals, and legal-adjacent copy?
Most brand voice rules assume the news is fine. An apology, a "no," or a sentence a lawyer will read tests the ones that don't — and the voice block needs an explicit answer for what happens then, instead of leaving it to whoever's writing that day to guess. Without that answer, teams tend to swing to one of two bad defaults: keeping every rule active and sounding upbeat about an outage, or dropping the voice entirely and sounding like a form letter from a company with no personality at all. Neither is necessary once the exceptions are written down once.
20. Apology or bad-news copy
Brand voice block: [PASTE VOICE BLOCK HERE]
Situation: [WHAT WENT WRONG OR WHAT'S DELAYED]. Audience: [WHO].
Write the message. Keep the block's vocabulary and point-of-view rules. Override any rule that assumes good news (upbeat openers, exclamation points, promotional asides) — this message leads with what happened, states what we're doing about it, and does not minimize the impact on the reader.
21. Refusal or "we can't help with this"
Brand voice block: [PASTE VOICE BLOCK HERE]
Request we need to decline: [WHAT THE CUSTOMER ASKED FOR AND WHY WE CAN'T DO IT].
Write the response. Keep our vocabulary and point-of-view rules. State the decline in the first two sentences rather than burying it after a long preamble — our voice values directness even here, where it would be easy to default to over-softening.
22. Legal-adjacent or compliance copy
Brand voice block: [PASTE VOICE BLOCK HERE]
Copy to write: [e.g. a data-retention notice, a refund-policy explanation, a terms summary for customers].
Keep our vocabulary and point-of-view rules wherever they don't conflict with accuracy. Where a legal requirement conflicts with a stylistic rule (e.g. a required disclosure phrase versus our "avoid corporate nouns" rule), keep the required phrase exactly and flag the conflict rather than silently rewording it. This is not legal advice, and the output should still be checked by whoever owns compliance.
23. Deciding when the voice has to yield
Use this one before an edge case comes up, not after: it turns "the voice yields sometimes" into a specific, written decision instead of a judgment call made under deadline pressure.
Brand voice block: [PASTE VOICE BLOCK HERE]
For each of the following situations, decide whether our voice rules apply as written, apply with a named exception, or should be set aside entirely, and say which specific rule is at risk in each case: an apology, a security incident notice, a price increase, a refusal, a legal disclosure, a message to a very upset customer. Output as a short table: situation, decision, which rule is affected.
How do you manage brand voice across multiple brands or personas?
An agency running five client accounts, or a company with a founder voice, a support voice, and a product-docs voice, has a different problem: not "does this match our voice" but "which voice is even active right now." Running one long session across brands is where voices bleed into each other, usually in the direction of whichever brand you were writing for two prompts ago. The fix is less about the prompt content and more about discipline in how blocks get stored and named: one block per brand or persona, never merged into a single mega-block, and a habit of naming which one is active before the model writes a word.
24. Multiple brands, one operator
I manage AI writing for multiple brands. Voice block for the brand I'm working on right now: [PASTE ACTIVE BRAND'S VOICE BLOCK HERE]
Before writing anything, confirm back to me in one sentence which brand's voice block you're using, so I can catch it immediately if the wrong one got pasted. Then write: [TASK].
25. Multiple personas, one brand
Our brand has more than one voice depending on who's speaking: [NAME EACH PERSONA, e.g. "Founder — first person, opinionated, used in newsletter"; "Support — plain, calm, used in help replies"; "Product docs — neutral, instructional"].
Active persona for this task: [WHICH ONE]. Persona-specific rules: [RULES FOR THAT PERSONA, on top of the shared brand rules below].
Shared brand rules that apply regardless of persona: [PASTE SHARED VOICE BLOCK HERE]
Write: [TASK], using the active persona's rules layered on top of the shared brand rules.
How does Prompt Architects fit into this workflow?
None of the 25 prompts above need our product to work: paste any of them into ChatGPT, Claude, or Gemini and they run as written. What a saved, parameterized prompt is for is not retyping this list every time. A voice block you built once with prompt 8 is exactly the kind of thing worth saving to a prompt library rather than hunting for it in a doc before every draft.
Two features map directly onto what's above. The Tone Selector, on our Advanced and Team plans, is the dial for the register column in the format table (formal, conversational, or somewhere between), layered on top of a voice you've already fixed, rather than a substitute for having fixed it. Contexts is where the voice block itself lives once you've built it with the extraction and encoding prompts, so it travels with you across tools instead of getting retyped; our guide to building that context covers the storage side in more depth than fits here. For a team of more than one writer, the team prompt playbook covers keeping a shared block from drifting once more than one person can edit it.
Free accounts get 5 prompt enhancements a day, published on our FAQ page, enough to run the extraction and encoding prompts above without needing a paid plan first. Applying the voice at real volume across email, landing pages, and support replies is where a saved library starts to earn its keep; our email marketing prompts and landing page copy prompts are worth pairing with the format prompts above if either channel is where most of your volume lives.
Stop rewriting prompts. Start shipping.
Works with ChatGPT, Claude, Gemini, Grok, Midjourney, Ideogram, Veo3 & Kling. 5.0★ on the Chrome Web Store.
Create An AccountStart with one voice block, not all 25 prompts at once: run the extraction prompts against three to five samples, encode the result into rules with prompt 8, then audit your next real draft against it with prompt 16. The gap between the audit's pass count and 100% is the actual list of what to fix — a shorter, more specific list than anything an adjective would have given you.