How To Write With An LLM

A short essay by Thomas and Erin Ptacek (the byline carries both names) on using a language model while you write, without letting it write for you.

Their starting point is that readers notice machine wording (“Readers can detect LLM words in the parts per trillion”), so you have to write for yourself. A model is still useful as a copyeditor. You write the piece first, then hand it over to find the flaws.

Two rules carry the piece. Rule Number One: “You may not use a single word an LLM suggests to you.” Frontier models are very good at pleasing turns of phrase, and that is the problem: any phrase the model offers is off limits, even when it reads better than yours. They call this “intellectual personal protective equipment”. Rule Number Two: avoid encouragement. A model that praises a draft talks you into keeping first-draft instincts that needed a rethink. For a couple of years their prompts opened by claiming the writer was an editor screening someone else’s submission; the model then fitted its advice to that invented publication. The advice now is to forbid praise and stay suspicious of any that gets through.

What the model is good for is the tedious part: passive voice used too much or too little, nominalizations, repeated phrases, filler words (“very”, “really”, “actually”) and paragraphs that would read better in a different order. The workflow has three steps. Ask the model to spot problems. Rewrite each flagged passage yourself. Then show the original and the rewrite to a model that has not seen your editing, and ask which is better. They built a small workshopping tool around this (an editor with sidebar commentary and revision tracking).

The piece gives no evidence beyond the authors’ own practice. “Parts per trillion” is a figure of speech. The one study of machine “tells” checked for this garden (Graphite’s, covered by Fast Company) counts them across thousands of articles, and its authors say single phrases should not be used to judge who wrote something. The Ptaceks also don’t take all of the model’s advice: GPT5 told them the piece was 20% too long, and they left it as it was.