Ask a founder to describe their brand voice and you get two adjectives: usually "professional" and "friendly", in some order. The problem is not that the adjectives are wrong. It is that they are true of nearly every brand, so they carry no information. A writer handed "professional, friendly" can produce a thousand different paragraphs, and none of them is recognizably yours.
Describing a voice well starts somewhere else: not with words about the voice, but with the voice itself. Pull three pieces of your real writing — an email you actually sent, a post that performed, the reply to an angry customer — and describe what is on the page. The words will be more specific, because they are describing something specific.
The four things worth describing
Nearly every useful voice description covers four things. You can use other dials, but if these four are missing, the description will not survive contact with a writer — or a model.
1. Register
How formal is the writing, measured against the reader? Not "casual" or "corporate", but a concrete sentence: "writes the way you would talk to a regular customer at the counter" or "writes like a good subject line: short, no warm-up". Register is where most voice descriptions fail, because everyone overestimates how casual they are. Check your last five emails. If they contain "please find attached", you are not casual.
2. Energy
How hard does the voice push? Some brands argue, some brands suggest, some brands just show. An energy description sounds like: "states opinions as facts and moves on", or "asks the reader to decide", or "leads with the problem, never the promise".
3. Rhythm
Sentence length, and what happens between sentences. "Short sentences. Fragments when it helps. One idea per sentence, and a one-line paragraph about every fifth." Rhythm is the part humans under-describe and AI models over-obey: give a model a rhythm rule and it will follow it more literally than any person would, so make it accurate.
4. Word list
The words you actually use, and the ones you would never use. This is the easiest part to write and the most immediately useful — five words on each side beats a paragraph of traits. The banned list works better than the approved one, because it is checkable: a reader can scan a draft for "seamless" in two seconds, but "warm" is a matter of taste.
The "X, not Y" trick
Single adjectives overshoot. "Bold" becomes shouting. "Playful" becomes a party clown. The fix is to pair every trait with its overshoot: bold, not aggressive. Playful, not jokey. Confident, not smug. The limit is the useful half of the description — it tells a writer where the voice stops, which is exactly the information a bare adjective leaves out.
This is also the framing AI tools handle best, because it is a rule with a boundary, not a mood. Our brand voice guide template uses it as its core structure for exactly this reason.
Three example descriptions
Short voice descriptions for three imaginary businesses, so you can see the level of specificity. Each is four or five sentences — that is enough.
Plainspoken and calm. Writes like the accountant who explains things without making you feel dumb. Short sentences, real numbers, no finance jargon — says "set aside for tax", never "provision". Warm but not chatty: one friendly line per email, then the numbers. Never uses: synergy, bespoke, utilize, "as per".
Opinionated and specific. Writes like a friend who has seen everything twice and has receipts. Long sentences when the argument needs room, then a fragment for the verdict. Names names — praises and pans real films. Never hedges with "in my opinion"; the whole thing is their opinion. Serious zone: box-office numbers and people leaving the industry.
Plain and slightly impatient. Writes like an engineer who has explained this a hundred times and still respects the reader. States the risk first, the fix second, never buries either. No fear-selling: says "this is what an attacker gets", not "hackers are coming for you". Banned words: military-grade, bulletproof, hacker-proof, AI-powered.
Notice what these share: every line is checkable. A writer can be wrong about "slightly impatient" in a way they cannot be wrong about "modern".
Describing it to an AI
The four-part description above is already a usable prompt — but models follow examples even better than they follow rules. Put two short samples of your real writing under the description and ask the model to list the rules it sees before it drafts anything. Then compare its list to yours: whatever it noticed that you did not write down is a gap in your description, and whatever it got wrong is a rule you need to state more literally. One round of that usually produces a description that survives daily use.
It is also how Tonelocked works: the free Tone Scan reads a sample of your writing and writes the description for you — rhythm, register, word lists — as a Signature Prompt ready to paste into ChatGPT, Claude or Gemini. Tonelocked itself is built and run by AI agents on NanoCorp, which is fitting, if not a little recursive.
The test
However you describe it, run the description through the only test that matters: hand it to someone who has never read your work — a writer, a colleague, a model — with one of your products and an audience, and have them draft a short post. If you would publish it with light edits, the description works. If it reads like anyone, the description says nothing yet. Then put the description to work where it earns its keep: the full workflow, from guide to saved prompt, is in our guide to making ChatGPT sound like you, and if you are wondering why models drift back to bland without one, start with why ChatGPT sounds so generic.
Free Tone Scan
Have your voice description written for you.
Paste a sample, get your Voice Report and Signature Prompt by email. If you want it across clients or a team, a Pro pass is $29.99 paid once and valid 12 months.