The Core of LLM Writing: Draft First, Edit Second
The safest approach to LLM writing is not to ask a model to generate an article from scratch. Write the first draft yourself, then use the model as a strict copyeditor. It can flag repetition, passive voice, nominalized language, unnecessary adverbs, and paragraphs that would work better elsewhere. The author should still control the judgment, ideas, and wording.
Two Rules for Protecting the Author’s Voice
The article proposes two deliberately strict rules. First, do not directly adopt a specific phrase suggested by the model. Models are good at producing smooth, striking language, but that can make an entire article sound like a sequence of headlines. Second, do not let the model offer generic encouragement. Early drafts often need structural changes, cuts, and rethinking—not reassurance that everything already works.
Keep Models on Mechanical, Valuable Work
A more useful workflow breaks editing into passes: first ask the model to identify problems, then review sentences, paragraphs, and structure separately. When comparing revisions, use a model that does not know which version the author just changed, reducing the pressure to praise the latest edit. The author also recommends maintaining a personal checklist of prompts and using software to track highlights, suggestions, and revisions so repetitive work can be delegated.
Its limits matter just as much. Model suggestions are not automatically correct and may make prose longer, smoother, or less distinctive. Models are best suited to tiring, rule-based checks, not to deciding what the author means. The open question is whether writers can gain speed while resisting the model’s phrasing and praise. If they cannot, editing assistance may quietly become ghostwriting.