The sentences are even. The rhythm is predictable. Every paragraph tries to sound important. Too many em dashes, tidy lists of three, metaphors doing overtime.
The text can still be fine. It just sounds like a machine. Once a reader starts spotting the pattern, the message stops landing.
Claude-speak is not a taste issue
Mason Hall wrote this week that he is tired of so-called Claude-speak. Nous Research answered with a practical fix: try the Hermes Humanizer skill.
Short exchange, right point. The model can write. It just writes the same way every time.
A bigger model will not fix the voice. Voice is an editorial choice. What to leave out, when to say the thing directly, whose name the message goes out under.
Do not ask it to sound more human
The default style is even and polite. A company voice is messier. Familiar words, short sentences, sometimes too blunt.
Asking it to "write more human" does not help. It swaps one mannerism for another. Warm metaphors show up, the transitions get softer, and the new pattern is just as easy to spot.
The customer notices first
A reply can be correct and still feel like a mass message. The first one may pass. By the third, the reader is already waiting for the next perfect list.
Trust does not fail on a wrong fact. It thins out because the message does not sound like anyone. Same on social: if every post sounds like the same model, people do not remember you. They remember the style.
Give samples, not a personality essay
A few real samples beat a long description. An email or a customer reply tells it more than "be relaxed but expert".
Put the words you hate on a blocklist. If you do not want em dashes, "unlock", or "a new era", cut them. Ask the agent to edit the draft before publishing. The first version is where the default voice is loudest.
Check that it does not invent opinions or certainty it has not earned. The most revealing AI line is not always the style. It is the sentence where the model sounds seasoned without having any right to.
We spent years teaching AI to produce more text. Now it has to learn what to leave out.