The diagnosis: adjective inputs produce adjective outputs. Three structural moves that change the result, and why a voice guide is for a person, not a brand.
You've seen the output before, even if it wasn't yours. "Direct, confident, approachable. Clear, concise language. Professional but warm. Conversational quality that builds trust." Open any AI-extracted brand voice and you'll find roughly that paragraph, with the adjectives shuffled.
The reason it's nearly identical across brands isn't that all brands actually have the same voice. It's that the prompts asking for the voice all share the same structural defect.
A voice guide is supposed to be portable. A new copywriter joins, they read the guide, they write copy that sounds like the brand. Adjective lists don't transfer because "direct" means something different to every writer. The output looks like a guide. It acts like a horoscope.
The team gets a document. Nobody's downstream work changes. The voice exercise produced a deliverable instead of an outcome.
When you ask Claude "what's the voice in these samples," the model has to pick between a token-level answer (adjectives) and a pattern-level answer (sentence structure, vocab habits, opener tendencies). The model defaults to the token level because that's what the question literally asks for. "What's the voice" resolves to "voice gets described in adjectives" resolves to a paragraph of adjectives.
The model isn't lazy. It's answering the question you asked.
The fix isn't "ask for better adjectives." Better adjectives still produce a horoscope. The fix is structural. Three moves change what the output is.
Don't ask for "three voice attributes." Ask for "three voice attributes, each paired with its opposite." The pairing is the lever.
"Direct" alone is generic. "Direct, not blunt" is specific. The opposite forces the model to commit to a positioning, not just a description. You can't write "warm, not cold" without telling me whether the brand prefers humor or sincerity. The pair extracts that decision and writes it down where a copywriter can see it.
A guide built on paired contrast tells you what the brand is by telling you what it deliberately isn't. That's the difference between "describe the voice" and "place the voice on a map."
Don't ask "what's the tone." Ask:
Patterns are reproducible. Adjectives aren't. A new writer can read "73% of paragraphs open with a claim, 14% with a hedge" and produce a paragraph that does that. They cannot read "direct" and produce something measurably more direct.
The pattern is also auditable. You can run the analysis on the writer's draft and see whether the new copy hits the spec. The adjective version offers no such test.
Don't ask the model to imagine what good and bad look like. Make it work from the source.
"Pull six examples from the samples where the brand uses the voice well. Then rewrite each one in the most common alternative phrasing. The contrast IS the guide."
The do/don't pair is more useful than the adjective because it shows the writer the substitution they need to make. "We help teams unlock the power of AI" becomes "Two teams are using it to replace their reporting analyst. The gap between those two uses is the whole opportunity." Same idea, different voice. The writer learns the swap. They can run the swap on their own draft tomorrow morning.
A voice guide is for a person. Usually a copywriter you trust who needs to ramp up on a new brand. Sometimes a Claude Project that needs to inherit the voice for everything that comes after. Often you, on a Tuesday morning, trying to draft an email that sounds like the brand instead of like an LLM that's been fed the brand.
Either way, the user is the protagonist. The guide is a working document a real human picks up to make better decisions. Not a brand asset. Not a slide for a stakeholder review. A tool a writer uses to do work they'd be doing anyway, slightly better.
The adjective-shaped guide fails that test. The pattern-shaped guide passes it. People are the whole point of the exercise. The prompt either serves them or it doesn't.
Here's how you know whether your voice extraction worked: hand the guide to someone else, give them a writing task, and watch what happens.
If they produce copy that sounds like the brand using only the guide as input, the prompt worked. If they have to come back to ask "what does direct mean to you," the guide is a horoscope. Cut it and run the structural version.
Two paths from here:
Either way, the test is downstream. The guide isn't the deliverable. The next piece of copy a human writes from it is.
Built around the three structural moves above. Three voice attributes paired with opposites. Pattern-level rhythm and vocabulary specs. Do/don't pairs from your real samples. CC BY 4.0, no email gate.
Get the prompt →