The moment a team starts generating social posts with AI, a new problem appears: the posts are grammatically perfect and completely faceless. They could belong to any SaaS company, any agency, any consultant. Brand voice is what separates your output from that grey middle, and an AI model will only protect it if you encode the voice in a form the model can read and a reviewer can score.
This guide is about the encoding and the enforcement, not the philosophy. It assumes you already have opinions about how you sound and need to make a language model honor them at scale.
Why models drift toward generic
Large language models are trained to produce the statistically average sentence. Left alone, they reach for “Unlock the power of,” “In today’s fast-paced world,” and “Let’s dive in.” Those phrases are the default because they appear everywhere. Your voice is, by definition, not average, so without explicit constraints the model regresses to the mean on every draft.
Three forces pull AI posts toward sameness:
- Vague instructions. “Make it friendly and professional” describes 90% of all brands and gives the model nothing to act on.
- No negative examples. Telling a model what to write is weaker than also telling it what you never write.
- Lost context between drafts. Each generation starts fresh unless the voice definition travels with every prompt.
Encode voice as data, not adjectives
A usable voice definition for AI has four parts, and each one should be concrete enough to test a draft against.
| Component | Weak version | Encodable version |
|---|---|---|
| Tone axes | “Confident but approachable” | Formality 3/5, warmth 4/5, humor 2/5, sentence length short |
| Lexicon | “Use our brand words” | Always: ship, build, team. Never: leverage, synergy, solutions |
| Sentence rules | “Be clear” | Max 22 words. No semicolons. Active voice. Lead with the verb |
| Gold examples | “Sound like us” | 8 real posts you would publish unedited, 4 rewrites of off-voice drafts |
The tone axes matter most for AI because they convert taste into numbers a prompt can reference and a reviewer can check. “Humor 2/5” means one light line is fine, a meme caption is not. The banned-phrase list is the single highest-leverage item: a model that knows “never write leverage, seamless, game-changer, in today’s landscape” instantly stops producing the worst 20% of generic output.
Put the voice into every prompt
A voice file that lives in a shared doc protects nothing. It has to be injected into the generation step. A practical prompt structure looks like this:
You write social posts for [company]. Voice: formality 3, warmth 4, humor 2. Always use these words where natural: ship, build, team. Never use: leverage, seamless, game-changer, dive in, unlock. Sentences under 22 words, active voice. Match the rhythm of these three examples: [paste 3 gold posts]. Draft a LinkedIn post from this source: [paste source].
The gold examples do more work than the rules. A model imitates patterns it can see far better than rules it has to interpret, so three pasted on-voice posts will fix more drift than three paragraphs of description. This is also where AI social media prompts and a maintained social media prompt library earn their keep: the voice block becomes a reusable header on every prompt.
Voice does not mean identical across channels
Enforcing voice is not flattening it. The same brand can be 3/5 formal on LinkedIn and 2/5 on TikTok and still be unmistakably itself, because the lexicon and the banned list stay constant while the tone axes flex. A useful rule: the words and the values are fixed, the register moves. A founder writing a sharp LinkedIn argument and the same founder writing a fast Instagram hook should share vocabulary and point of view even when sentence length changes. If you run founder-led social media , capture that person’s real phrasing as gold examples rather than inventing a corporate voice they will never sound like.
Enforce voice with a score, not a vibe
Reviewers catch off-voice drafts faster when they grade against the same axes the prompt used. A lightweight voice score before publishing:
- Read the draft aloud. If it sounds like a press release and you wanted warmth 4, it fails.
- Scan for any banned phrase. One hit is an automatic revise.
- Check the lexicon. Did at least one brand word appear naturally? Forced insertion also fails.
- Compare against the closest gold example. Same rhythm and confidence, or noticeably blander?
Track the revision rate for voice reasons over a few weeks. If 40% of drafts get sent back for tone, your prompt is under-specified, not your team. Tighten the banned list and add the corrected drafts as new negative examples. Voice enforcement is a loop: every off-voice post the model produces is a missing example you can feed back in. Keep this work close to your broader social media content quality standards and your social media brand guidelines , which define the wider rules this voice work sits inside.
Where Utin fits
Utin is being built so the voice definition travels with the work automatically. It reads your website to learn how you already write, carries tone axes and a banned-phrase list into every generated draft, and flags off-voice output before it reaches a reviewer. If you want voice enforcement that improves as the model sees more of your corrections, you can register interest in an early pilot.