Almost every company that has ever hired an agency owns a file called “Tone of Voice”. Fourteen slides, the words “friendly”, “professional” and “innovative”, two stock photos. It sits in a shared drive, last opened the day the project was signed off. Meanwhile the posts get written however: warm and full of emoji when Maya writes them, “we hereby announce the commencement of our sale” when the new intern does.
That was survivable when a single editor read everything. It stopped being survivable now that drafts come out of a model. With no voice defined, AI defaults to the average voice of the internet: “in today’s fast-paced world”, “it’s not just a tool, it’s your trusted partner”, three emoji at the start of the line and an exclamation mark at the end. Readers recognise it in two seconds and stop reading.
Below: how to build a usable tone of voice in one evening, fit it on a single page, hand it to a model so it actually holds, and a 60-second edit checklist that catches machine phrasing.
Voice and tone are different things
Voice is character and it doesn’t change: who you are to the reader, how formal you are, whether you joke. Tone is how that character sounds in a specific situation. One voice, several tones:
| Situation | Tone | Sample line |
|---|---|---|
| Feature announcement | Calm, factual | ”Drafts from RSS are live. One click to turn them on” |
| Outage | Short, no excuses | ”Instagram publishing was down 14:10–15:40 UTC. Expired token on our side. Fixed” |
| Angry comment | Level, no irony | ”Fair point, that is annoying. Here’s what we can do right now: …” |
| Teaching post | Patient, with examples | ”Start with three slots a week. Why not seven — below” |
| Selling | Direct, no pressure | ”$29 a month, covers five accounts” |
If your document describes only the voice, the team jokes in the same register during a launch and during an outage. That is where half of all social media apologies come from.
Four sliders instead of ten adjectives
Adjectives don’t work because “friendly and professional” is a description every brand would sign. Sliders work because they force a choice. Put your own numbers from 1 to 5 on each:
- Formality. 1 is “hey, look what we found”, 5 is “Dear valued partners”.
- Humour. 1 is none at all, 5 is memes and self-deprecation as house style.
- Push. 1 is “details here if you’re curious”, 5 is “grab a seat before they’re gone”.
- Depth. 1 is beginner-level advice, 5 is numbers, API limits and edge cases.
Our own profile, as an example: formality 2, humour 2, push 2, depth 4. Those four numbers already settle most arguments. Someone drafts a snarky post about a competitor: humour 2 and push 2 say no. Someone asks how many posts a day Instagram allows: depth 4 demands the actual limit, not “it depends on your account”.
Finding the voice in one evening
- Pull ten of your own posts you actually like. Not the top performers — the ones you’d sign your name to. Write down what repeats: paragraph length, how you address people, whether there are questions, where the jokes sit.
- Record a voice note. Have the founder, or whoever explains the product best, talk into a phone about why any of this matters. The transcript is the most honest first draft of your voice. Move the good phrases into the document verbatim.
- Pick two anti-examples. Two brands whose tone is wrong for you, with the reason: “too much caps”, “world-changing grandeur for a to-do app”. Boundaries define a voice more precisely than descriptions.
- Build a 10/10 vocabulary. A “we say / we don’t say” table: “tool” not “platform”, “the charge goes through” not “the transaction will be processed”, “outage” not “service degradation event”.
- Run the recognition test. Strip the logo off three posts and show them to a colleague mixed in with competitors’ posts. If they can’t tell which is yours, you don’t have a voice yet.
The one-page document
Nobody reads a long one — not humans, not models. The working format is a single page:
- Who you are to the reader: peer, mentor, expert, seller.
- The four sliders, with numbers.
- Five rules written as permissions and bans (“no caps”, “numbers are always specific”, “we never say ‘simply’”).
- Vocabulary: ten words in, ten words out.
- Three full reference posts, each with a note on why it works.
- Three situational tones: outage, sale, criticism.
Keep that page where people see it — in the brand knowledge base, next to the content plan, not in a folder called “Branding 2024”.
Handing the voice to AI
Order matters more than length. What actually moves the output:
- Examples beat descriptions. Three to five reference posts in the prompt do more than a page of adjectives. Models pick up rhythm, sentence length and vocabulary from samples, and almost nothing from abstractions like “write in a friendly yet professional manner”.
- A stop-list is mandatory. Explicit bans work: “never open with ‘in today’s world’”, “no ‘it’s not just X, it’s Y’”, “one emoji maximum, never at the start of a line”, “no percentages unless they come from our own numbers”.
- Keep facts separate from style. Voice governs form; facts govern content. Prices, limits and plan names should never be invented — they come from your knowledge base. The mechanics are in our piece on AI posts that don’t sound like AI.
- Ask for three variants, not one. Choosing between three is faster than rewriting one. It also exposes the pattern the model keeps repeating, which tells you which example to add.
- Store the voice in a system, not in a chat. Threads get lost; documents don’t. In Repost21 the brand knowledge base connects live to Google Docs or Notion: you edit the doc and every following draft changes, with no prompt surgery.
The stop-list: what gives AI away
- Warm-up openers: “in today’s fast-paced world”, “let’s dive in”, “it’s important to remember”.
- “It’s not just X, it’s Y” and its cousin “not only… but also”.
- Symmetrical triads: three adjectives, three bullets, all the same length.
- Even rhythm: every paragraph four lines, every sentence mid-length. Human writing lurches.
- Emoji used as formatting — a smiley starting every bullet.
- False precision: “studies show 73% of companies…” with no source. Invented numbers are the fastest way to lose a reader; if it’s your own measurement, say so.
- A closing paragraph that summarises what you just read.
The 60-second edit: read the first line (hook or warm-up?), find one vague phrase and replace it with a number, delete half the emoji, read the last paragraph out loud. If it sounds like a toast at a wedding, rewrite it.
One voice, different networks and languages
The character stays, the delivery changes: shorter and blunter on X, drier and more numeric on LinkedIn, conversational on Threads, like a message to a colleague in Telegram. That’s one voice with per-network settings, the same logic as content repurposing.
Languages are harder. A tone of voice doesn’t transfer word for word: formality works differently in Spanish and Uzbek, and jokes rarely survive translation. The practical rule is to set sliders and vocabulary per language, and to write the reference posts natively rather than translating them. Otherwise your second language reads like a translation — which, to a reader, is the same signal as “a machine wrote this”.
When tone of voice won’t help
- You have nothing to say. A voice doesn’t replace an offer, and it won’t rescue a weak product. Selling still needs structure — see posts that sell.
- The document was written by an agency but the founder does the talking. The founder wins, which means the doc should have started from them.
- The voice changes every quarter. Recognition is built by repetition; a rebrand every three months resets it to zero.
- One voice stretched over different audiences. An enterprise sales motion and a teen product line will not share a slider profile. That’s two documents.
- The voice exists only in posts. If DMs get answered with canned scripts, readers read the voice as a marketing mask.
A brand voice is cheaper to configure once than to re-impose on every draft. In Repost21 it lives in the knowledge base, Luna writes per-network variants from it, and prices and facts come from your own documents — setup details are in the documentation, and you can test it on your own posts on the free plan.