Use an LLM to edit your writing, never to write it

Thomas Ptacek argues that an LLM is a good copyeditor and a bad ghostwriter, and the way to keep the benefit without the cost is one hard rule: you may not use a single word the model suggests. His reasoning is that frontier models are very good at producing pleasant phrases, and the moment you paste one in, your prose starts to smell like every other piece of AI text, flat and shaped like a headline. So you write the draft yourself, then use the model only to find mechanical problems: passive voice, filler words like "very" and "actually", clumsy paragraph order.

He adds a second rule that is easy to miss: turn off the encouragement. Models reflexively praise a first draft, and that praise quietly discourages the rewriting that real editing requires. Ptacek suggests a small workflow instead. Have the model flag problems, rewrite them yourself, then compare the two versions using a fresh model instance so its earlier suggestions do not bias the judgment. He built his own tool in Python and HTMX to run these passes and shared the system prompt.

Simon Willison, who links the essay, follows a similar practice: he uses custom proofreading prompts but refuses to let a model generate the actual content. Both treat the line between editing and writing as the thing to protect. The tedious mechanical work goes to the machine; the voice stays yours.

Why it matters

If you write for work and worry that AI is eroding your voice, this is a concrete discipline you can adopt today: draft it yourself, let the model catch errors, and reject every phrase it hands you. It keeps the speed without the sameness that readers now notice and distrust.