How to write with an LLM: copyediting only, never use its suggested phrasing
Simon Willison · rss · 2026-09-18
Security expert Thomas Ptacek proposes using LLMs as copyeditors, not writing assistants, with one strict rule: you may not use a single word the LLM suggests.
- LLM-generated text has a distinctive "weird smell"; banning suggested phrasing keeps your own voice and enforces discipline.
- Simon Willison endorses this: he uses LLMs for fact-checking, grammar and as a thesaurus, but never for blog content, and shares his proofreading prompt.
- The piece includes a screenshot of Ptacek's personal LLM copyediting tool and a prompt to kickstart building your own.
More from Apps
- Asking Codex to explain papers by replacing all math with PyTorch notation — nabla_theta · 2026-09-18
- New Pika launches as all-in-one AI creative platform that handles model plumbing — Kyrannio · 2026-09-18
- OpenWand ships MIT-licensed LLM interface with one-click context fetching and model freedom — Sunnyli1337 · 2026-09-18
- Selling AI Kitchen Renovations: GPT-6 Astra + Higgsfield API Pitched as a $10k/mo Business — eptwts · 2026-09-18
- An AI agent filed a dependent care FSA claim end to end, pulling receipts from Gmail — armand_ruiz · 2026-09-18
- Indie app EchoVault gives its Gemini-powered digital clone a 3D brain visualizer — Emojinapp · 2026-09-18