Prompting for Brand Voice & Creative Work
Capturing a specific, consistent voice across dozens of pieces of content
Voice consistency is a systems problem, not a one-off prompt problem
Getting Claude to nail a brand voice once is easy — show it a couple of good examples and ask for more of the same. Getting it to nail that voice consistently across the fiftieth piece of content, written by someone else on your team, in a different conversation, weeks later, is a systems problem. This lesson is about building a voice system that survives beyond a single good prompt.
A voice reference isn’t a list of adjectives
“Friendly, confident, and clear” describes almost every brand and constrains almost nothing. A usable voice reference is built from concrete examples and explicit contrasts, not adjectives.
BRAND VOICE REFERENCE
We sound like: a sharp colleague explaining something over coffee,
not a marketing department.
DO:
"We tested this with 40 customers before shipping it."
"Here's the part that usually trips people up."
DON'T:
"We're thrilled to announce this exciting new feature."
"Leverage our best-in-class solution to unlock your potential."
Sentence length: short. If a sentence needs a comma to hold two
ideas, it's probably two sentences.
Humor: dry, occasional, never at the customer's expense.
Never use: "game-changer," "unlock," "seamless," "elevate,"
exclamation points in body copy.Concrete do/don’t pairs, a banned word list, and a stated rhythm rule (sentence length) do more work than any adjective list, because they’re directly checkable — you can look at a sentence and tell whether it violates the rule.
Calibration prompts before production prompts
Before generating real content, run a calibration pass: give Claude the voice reference and a neutral piece of text, and ask it to rewrite that text in voice. Check the result against your own ear before using the reference for actual production work. This catches ambiguities in your voice reference early, where they’re cheap to fix, rather than after fifty pieces of content have gone out slightly off.
Key principle: If two people using the same voice reference produce noticeably different content, the reference is still too abstract. Keep adding concrete do/don’t examples until the outputs converge.
Anchoring with your own best examples, not generic ones
The strongest voice references are built from your own highest-performing real content — the email that got the best reply rate, the post that actually sounded like your founder. Generic “brand voice” examples from a style guide template rarely capture what’s actually distinctive about how your brand sounds versus every other brand claiming to be “friendly and confident.”
Consistency across content types
A voice that works for a product email may need slightly different calibration for a LinkedIn post or a support message — same core voice, different register. Build one core voice reference, then a short addendum per content type noting what shifts.
| Content type | What shifts from the core voice |
|---|---|
| Support replies | Warmer, more direct empathy, still no exclamation points |
| LinkedIn posts | Slightly more personality, first person, can be more casual |
| Product docs | Same clarity rules, but zero humor, purely functional |
Auditing drift over time
Voice drifts. Periodically run a self-critique pass (Lesson 6) specifically checking recent content against the voice reference’s do/don’t list and banned words — this catches slow drift before it becomes a noticeably inconsistent body of content.
Check this batch of 5 recent posts against our voice reference.
Flag any sentence that uses a banned word, exceeds the sentence
length rule, or reads like generic marketing copy rather than our
voice. Quote the specific offending sentence for each flag.Build it: Take your current brand voice description. If it’s mostly adjectives, rewrite it as 3-5 do/don’t sentence pairs pulled from your actual best content, plus a banned word list.