How to Use LLMs Without Losing Your Writing Voice

Summary

The author argues that large language models are most useful in writing when treated as copyeditors rather than ghostwriters. Writers should produce their own drafts first, because model-generated phrasing can sound polished but make a piece feel like machine output. The first rule is absolute: do not use any specific word or phrase suggested by the model. The author says frontier models are unusually good at producing appealing, headline-like language, and that even attractive suggestions can gradually flatten an individual voice. The second rule is to avoid encouragement. Models tend to praise drafts and reinforce the author’s initial structural choices, even when paragraphs are weak, the argument is poorly organized, or substantial material should be removed. Instead, the writer should remain responsible for rethinking and replacing the draft, while asking the model to identify mechanical problems such as passive voice, nominalizations, repeated wording, unnecessary intensifiers, and misplaced paragraphs. The proposed workflow uses a prepared set of editing prompts in multiple passes, asks the model to identify problems and generate alternatives for comparison, and has a separate model judge competing versions without knowledge of the editing history. The author also describes building a Python, HTMX, SQLite, and Tailwind workshopping tool that supports highlighted commentary, revision tracking, and CLI-based model runs. The final qualification is that writers should not accept all model advice: the tool can make editing faster and less exhausting, but preserving the work’s voice still requires human judgment.