How to Use LLMs as a Copyeditor, Not a Ghostwriter
Original: How to Write with an LLM
Why This Matters
A rare framework for preserving authorial voice while still extracting genuine utility from LLMs.
Writer Thomas Ptacek argues LLMs should be used strictly as copyeditors, not ghostwriters. His two rules: never use a single word an LLM suggests, and block all encouragement from the model. Readers detect AI phrasing at trace levels, so your draft must start—and stay—yours.
Ptacek's core argument is blunt: readers detect LLM-generated prose even when it's been heavily edited, and no amount of 'humanizing' reliably fixes it. His workflow has two steps—write the piece yourself first, then feed it to a model to surface flaws. The method hinges on two strict rules.
Rule One: never use a single word the model suggests. Frontier models are exceptionally good at generating pleasing phrases, which is precisely the problem. Every suggestion tends toward magazine-headline register. Even when a suggestion feels like an improvement, Ptacek says treat it as disqualified—because you won't consistently spot the subtle ways a model will 'turn your writing into Velveeta.'
Rule Two: eliminate encouragement. LLMs default to telling you your draft is strong, your structure is solid, your metaphors land. That validation causes writers to double down on first-draft instincts rather than rethink them—and it's exactly that rethinking, Ptacek argues, that carries your voice. He previously tricked models by pretending to be a publication editor rather than the author, which helped but caused the model to overfit to imagined editorial goals. His current fix: explicitly forbid the model from offering praise, then stay skeptical of any that slips through.
What LLMs are genuinely useful for, in his view, is mechanical flaw-detection—spotting logic gaps, unclear transitions, redundant words—the tedious work a good human copyeditor would do.