Content operations
Voice Without the Cringe
Every tone-of-voice document ever written says "professional but approachable." No model has ever been helped by it, because it describes nothing.
Key takeaways
- Adjectives are unusable; encode voice as constraints, vocabulary, and structure.
- The banned list does more work than the aspirational list.
- Sentence-length variance is the strongest single signal of human authorship.
- Extract voice from your best existing writing rather than inventing it.
Ask ten people to write in a voice that is "confident, warm, and expert" and you will get ten different documents. Ask a model and you will get the average of everything it has read that was labelled that way, which is to say, the exact house style of every B2B newsletter you have ever unsubscribed from.
Voice is not a set of adjectives. It is a set of decisions, and decisions can be written down.
Extract, do not invent
Take five pieces you wrote that felt like you and performed well. Then answer, with specifics:
- Opening move. Do you start with a claim, a scene, a number, or an objection? Pick the one you actually use.
- Sentence rhythm. Measure it. Mine runs 8-32 words with a mean near 18. That variance is the signature.
- Vocabulary you own. The eight to twelve terms that recur because you think in them.
- Vocabulary you refuse. Longer list. More important.
- Evidence habit. First-party numbers, customer anecdotes, or cited research. Your default source of authority.
- Attitude to the reader. Peer, teacher, or provocateur. Only one, consistently.
Six answers, all concrete, all checkable by a critic model. That is a usable profile. "Professional but approachable" is not.
The banned list carries the weight
Positive instructions nudge; prohibitions bind. The constructions worth banning outright:
- "It is not just X, it is Y". The most recognisable AI construction currently in circulation.
- "In today’s fast-paced landscape" and every variant of the throat-clearing opener.
- "Let us dive in," "unlock," "leverage," "seamless," "robust," "game-changer."
- Three-item lists where all three items have identical grammatical shape. Real thinking is lumpier.
- Rhetorical questions used as section transitions. Sound familiar? Exactly.
- The tidy resolution paragraph that restates everything already said.
Rhythm is the tell
The most reliable signal that a human wrote something is uneven sentence length. Models converge on 15-20 words, paragraph after paragraph, and the effect is a kind of pleasant droning that readers detect before they can articulate why.
So specify it as a measurable constraint: a paragraph should contain at least one sentence under ten words and at least one over twenty-five. It is a crude rule. It works.
Slop is rarely wrong. It is regular. Regularity is what the reader is detecting.
Voice cannot survive an empty brief
The limitation worth being honest about: no voice profile rescues a piece with nothing to say. Voice is how a position sounds. Without a position, a strong voice just makes the emptiness more audible, a confident, rhythmically varied, unmistakably branded piece of nothing.
Which is why the profile belongs downstream of the research stage, not as a substitute for it. See the four-agent newsroom and the anti-slop standard.
Maintain it like code
Version the profile. When a rejection reason is "not my voice," diff the draft against the rules and find which one failed to bind, usually the profile is silent on that case. Add the rule, note the date, and keep the file under a page. A three-page voice document is a document nobody, human or model, will follow.
The voice interview script I use with founders is at alexcinovoj.com (opens in a new tab), and the codified version sits inside the brand profiles we build at TechTide AI (opens in a new tab). Voice is captured, not invented.
“A voice profile made of adjectives produces adjectives. A voice profile made of rules produces writing.”
Questions people actually ask
- How do I train AI to write in my voice?
- Encode decisions rather than adjectives: opening move, sentence-length range, owned vocabulary, banned constructions, evidence habit, and stance toward the reader. Extract these from writing you already like rather than describing an aspiration.
- Do few-shot examples work better than rules?
- They work together. Examples convey rhythm well; explicit bans catch the failures examples miss. Examples alone tend to produce pastiche of the specific samples.
- Why does AI writing feel generic even with a style guide?
- Because most style guides list adjectives, which are unverifiable. A critic can check for a banned phrase; it cannot check for "approachable."
- How long should a voice profile be?
- Under one page. Longer profiles compete with the actual brief for context and attention, and the marginal rules are usually the least enforceable ones.
Sources & further reading
Written by
Alex Cinovoj, Founder, TechTide AI
Alex Cinovoj builds ClawLI at TechTide AI, a fleet of specialist agents that research, draft, schedule, and grade LinkedIn content with a human approving every send. He writes about what breaks when you put agents in front of a real audience.
Keep reading
- Content operationsThe Four-Agent NewsroomA working blueprint for content operations: researcher, writer, critic, and scheduler, with contracts between them and a human at the gate.
- Content operationsApproval Is the ProductHuman-in-the-loop is usually sold as a safety compromise. Designed properly it is the fastest quality loop you have.
- Content operationsThe Anti-Slop StandardA publishable checklist for beating AI slop: what to ban, what to require, and how to enforce it before a draft ever reaches a human.