That was not a good attempt at changing the style.

You can't just say "don't sound like an LLM." The LLM does not in fact know that it is "speaking like an LLM"; it just thinks that it's speaking the way the "average person" speaks, according to everything it's ever been shown. If you told it "just speak like a human being"... that's what it already thought it was doing!

You have to tell the LLM a specific way to speak. Like directing an image generator to use a specific visual style.

You can say "ape the style of [some person who has a lot of public writing in the base model's web training corpus — Paul Graham, maybe?]". But that coverage will be spotty, and it's also questionably ethical (just like style-aping in image generation.)

But an LLM will do even better if you tell it to speak the in some "common mode" of speech: e.g. "an email from HR", or "a shitpost rant on Reddit" or "an article in a pop-science magazine."

Like “dead give away” instead of “dead giveaway”?