Thomas Ptacek published a discipline for working with LLMs that flips the default assumption. The tool checks facts, catches spelling errors, and acts as a thesaurus. It does not write. Rule Number One, in his words: "You may not use a single word an LLM suggests to you."

Simon Willison's Weblog surfaced the piece and endorses the framing. Willison applies the same boundary on his own blog, using LLMs for proofreading but not for content. Ptacek calls the restriction "intellectual personal protective equipment" and notes that AI-suggested phrasing carries a detectable quality he calls a "weird smell."

Ptacek also shared his personal copyediting tool as a screenshot and released his system prompt in a Hacker News comment. That makes the framework replicable. Watch whether this operator-grade discipline, use the model for process, own the output entirely, spreads as a professional norm rather than staying a niche writer's preference.