Use an LLM as a Copyeditor, Never a Ghostwriter

The core argument is that readers can sense machine-generated prose almost instantly, so you should still write everything yourself and only bring in a model afterward to catch flaws. The author lays out two hard rules. First, never adopt a specific word or phrase the model suggests—frontier models are unnervingly good at producing headline-grade turns of phrase, and stitching those into your draft pushes it into the uncanny valley between real expression and mere output. Second, block the model from praising your work, because its reflexive encouragement tricks you into keeping weak first-draft choices instead of doing the rethinking and rewriting that actually carry your voice.

Where models genuinely earn their keep is mechanical diagnosis: flagging overused passive voice, nominalized verbs, repeated phrasing, filler words like ‘very’ and ‘actually,’ and paragraphs that would read better relocated. The author frames this as tedious pattern-matching that a tireless model does better than a human editor, and recommends pairing it with a craft reference like Williams’ ‘Style: Lessons in Clarity and Grace’ to build a repeatable set of editing-pass prompts. One practical wrinkle: when comparing an original passage against your rewrite, hand the two versions to a model that lacks the context of your edit, or it will simply tell you the newer one is better.

The piece also describes a custom Python/HTMX/SQLite workshopping tool the author built to run these passes—highlighting text with Genius-style margin comments and dispatching prompts to CLI coding agents—while insisting that any tool you build for yourself will beat a borrowed one. The closing note captures the whole philosophy: GPT-5 told the author the essay ran 20% long, and he agreed it was probably right, then declined to cut it. Keep the labor-saving diagnosis; refuse the homogenization.