"Make it sound more human" is the most common tone instruction and close to the least effective one. It gives the model nothing to act on, so it does whatever it already thought human sounded like — which usually means inserting contractions and a rhetorical question.
Tone is not a mood that can be requested. It is the accumulated result of a small number of concrete decisions about how sentences are built: how long they run, which words are allowed, how confident the claims are, and how the paragraphs move. Those decisions can be stated, and once stated they are followed reliably.
The payoff is that tone stops being something you fix by rewriting, and starts being something you specify once.
Why the usual instruction fails
When you ask for a more human tone, the model has no target. It has a general sense of what "human" writing looks like across millions of examples, and that average is exactly the problem — it is the voice of nobody in particular.
The instruction also arrives too late. Tone requests usually come after a draft exists, which means the model rewrites at the sentence level while leaving the structure intact. But most of what makes writing feel mechanical is structural: uniform paragraph lengths, every section opening with the same construction, every claim landing at the same level of confidence.
Specifying tone before the draft avoids both problems, because the choices are made once and applied throughout.
Tone is a set of choices, not a feeling
Take a piece of writing you like and try to describe it without using adjectives. "Friendly" and "professional" are labels; they are not instructions. But you can describe the same piece in terms of decisions:
- Most sentences run between eight and fifteen words, with an occasional short one after a long one.
- Contractions are used freely in body text, never in headings.
- Direct address ("you") appears several times per section.
- Claims about results are stated plainly; claims about causes are hedged.
- Paragraphs are two to four sentences and rarely open with a transition word.
That description can be handed to a model, and it will follow it. The labels could not.
The four levers worth naming
Most tone variation that matters is produced by four things. Naming them explicitly in a brief is enough to control the output.
- Sentence length. State a working range and whether short sentences are used for emphasis. A piece where every sentence is 20 words reads as monotonous regardless of vocabulary.
- Vocabulary register. Which words are in and which are out. Banning a short list of overused terms ("leverage", "seamless", "robust", "delve") does more for perceived quality than any positive instruction.
- Confidence level. Distinguish between claims you assert and claims you qualify. A piece where every sentence is equally certain reads as either naive or insincere.
- Rhythm. The pattern of long and short across paragraphs. Deliberate variation is what makes prose feel written rather than generated.
Writing a style brief
A style brief is a short, reusable document — fifteen to thirty lines — that carries these decisions and travels with every prompt you run. It replaces the paragraph of adjectives people usually paste in.
Structure it in three blocks:
- Always. Constructions to use. Direct address, active voice, concrete examples in place of general statements.
- Never. The banned list. Specific words, specific openings, specific sentence shapes.
- Prefer. Decisions that hold most of the time but yield to clarity when they conflict. Ranges rather than exact values.
The "never" list is the highest-leverage part and it should be built empirically. Every time you delete a phrase from a draft, add it. After a month the list is a precise description of what your publication does not sound like, which is more useful than a description of what it does.
Editing for rhythm
Rhythm is the part models handle least well and the part readers notice most. Uniform sentence lengths produce the flatness people describe as "robotic", even when every sentence is correct.
One editing pass is enough. Read the draft and mark the length of each sentence. Where you find three or more consecutive sentences of similar length, break one of them into two, or merge two into a longer one with a subordinate clause.
The effect is immediate and disproportionate. A paragraph of fourteen, twelve, fifteen, thirteen-word sentences becomes fourteen, six, twenty-two, nine — the information is identical, and it reads as though someone wrote it.
Hedging and confidence
Generated text tends to hedge everything at the same level. "This can potentially help improve results in some cases" is technically safe and completely useless, because it commits to nothing.
Decide which claims your publication stands behind and state those plainly. Reserve hedging for the claims where the evidence genuinely is partial, and say why: "this works in most cases, but breaks down when the source material is under 500 words."
A precise qualification is more credible than a vague assertion, and it is more useful to the reader. The instruction to give the model is not "be more confident" — it is to name the conditions under which each claim does and does not hold.
Testing a style brief across samples
Do not judge a brief from one output. Run it on three different tasks — a short product description, a long explainer, and a list-style piece — and read them together.
Consistency is the thing to assess. If the three pieces sound like three different authors, the brief is ambiguous somewhere, and the ambiguity will compound as more people use it. If they sound like one author writing in three modes, the brief is doing its job.
Also read them aloud. Tone problems that are invisible on the screen — a repeated rhythm, a tic in sentence openings — become obvious when spoken, because speech does not allow skimming past the repetition.
Keeping tone stable over time
Store the brief next to the content it governs, version it, and treat changes as decisions rather than preferences. A brief that gets quietly edited every week produces output that drifts without anyone noticing, until someone compares a piece from March with one from September and finds they read like different publications.
Review it quarterly against your best-performing work. Whatever that work does that the brief does not describe is the next revision — and that is a much better source of updates than anyone's opinion about how the writing should sound.