Bots are given away by form rather than meaning: too even, too polite, too complete. All three are fixed in the prompt.
Length and unevenness
People write unevenly — sometimes one line, sometimes three. A model defaults to a paragraph of consistent length. An explicit instruction — «sometimes answer in a single sentence» — removes machine evenness better than anything else.
Ban the clichés
List what must not appear: «thanks for your question», «great question», «hope this helps», «feel free to reach out». Models avoid what is named, and those are exactly the phrases that read as automation.
Incompleteness
People do not answer exhaustively. They say the main thing and wait for a follow-up. «Do not try to close the topic in one message» makes the exchange feel live and extends the conversation.
Emoji
Either none, or one and not in every message. Three emoji in two lines is a reliable marker.
What does not help
Asking a model to «write like a human». That is a wish, not an instruction, and it produces nothing. Specific bans and specific requirements do.
And do not grow the prompt indefinitely: past a certain size the model loses priorities while every line adds to the per-call cost. Twenty precise lines beat two pages. The phrases to ban are listed in the giveaway phrases piece.