Voice transfer from adjectives doesn't work. Voice transfer from examples does. Here's the exact prompt, and the walkthrough.
"Help me write in our brand voice. We're warm, confident, and direct." If you've prompted Claude this way, you've probably noticed the output is fine, and entirely generic. It reads like every other warm, confident, direct brand. Because that's exactly what Claude's default for those adjectives is.
Adjective-based voice transfer doesn't work. Not in Claude, not in GPT, not in Gemini. Examples-based voice transfer does. This post is the why and the how.
Every prompt where you describe your voice in adjectives ("we're warm, confident, direct") wastes tokens because Claude's default output for those adjectives is the generic average. You haven't transferred voice. You've described temperature.
The fix is structural: extract once, paste the structured output into your Project, never describe in adjectives again.
Pretend Claude is a new senior copywriter you just hired. Two onboarding paths:
Path A: "We're warm, confident, and direct. Now write our next email."
Path B: "Here are five of our best emails from the last six months. Here are three we wrote that fell flat. Here's why each one worked or didn't. Now write the next one."
Path A produces generic output because adjectives don't have a single referent. "Warm" to one team means "uses contractions and the reader's first name." To another team it means "writes like a friend who knows their stuff." There are a hundred valid interpretations of "warm." Claude picks the statistical average, which is the boring middle of all of them.
Path B forces Claude to pattern-match on actual sentences. Sentence rhythm. Vocabulary. Where you break a paragraph. What you refuse to say. Voice is the residue of a thousand small choices. You can only transfer it via examples.
The Voice & Tone Extractor is one of the four free prompts in the library. It produces a structured voice guide you can paste into Claude Projects as your system prompt. Six sections:
"Direct, not blunt." "Specific, not granular." "Skeptical, not cynical." The opposite is the part most voice guides skip. Without naming what your voice is not, you've not constrained the model, you've just hinted in a direction.
Average sentence length (computed from at least 500 words of sample). Variance pattern. Em-dash usage per 300 words. Paragraph opening conventions. Adverb usage and what work adverbs do when they appear.
3-5 patterns the brand reaches for repeatedly. A particular metaphor family. A recurring transition. A structural move ("name the mistake first, then the fix").
15 words you actually use (drawn from the samples, not invented). 15 words you avoid, usually the obvious AI tells ("leverage," "unlock," "seamless," "transformative") plus your specific industry clichés.
Same idea, written two ways. The off-brand version is what a generic ghostwriter or default ChatGPT would produce. The on-brand version is what you'd actually publish.
A 60-80 word paragraph a new hire (human or AI) could read once and internalise. Distills the three attributes plus the most important signature moves.
Pick 5-20 pieces of your best on-voice brand writing. In rough order of value:
Avoid: anything ghostwritten by a freelancer, anything translated, anything older than 18 months. Those will pollute the average.
Open Claude. Paste in the Voice & Tone Extractor prompt (it's here, free). Paste your samples after it. Hit go.
Read the three voice attributes carefully. The "opposite" half is where the work happens, does the opposite feel off-brand? If not, push back on Claude and ask for a sharper opposite.
Look at the vocabulary lists. Are there words in "we use" that you don't actually use? Words in "we avoid" that you'd happily use? Edit by hand.
The do/don't pairs are where most teams catch issues. If the on-brand example doesn't sound like you, your samples were inconsistent. Either pick tighter samples or accept that your voice is more variable than you'd like.
This is where the leverage compounds. Take the output and paste it into your Claude Project's Instructions. Every future conversation in that Project inherits the voice without you having to remind it.
Voice drifts. The do/don't pairs especially go stale as the team grows. Every 90 days, re-run the Extractor on your latest 5 pieces. Update the Project Instructions.
Once the voice extraction lives in your Project, every prompt downstream inherits it. The content multiplier produces on-voice repurposing. The campaign brief comes back in your tone. The social posts sound like you wrote them.
The voice extraction is the single highest-leverage 30 minutes you can spend on your AI workflow. It compounds every prompt for the next 90 days. That's why it's the first move I'd recommend to any marketer setting up Claude, before the agent blueprints, before the playbooks, before any of it.
The full prompt, a real sample output, and the four-step walkthrough. Run it once, paste it into your Project, watch every future prompt inherit your voice.
Open the free prompt →