Improves content quality and brand voice by establishing safer workflows for AI-assisted writing.
How Do You Write With An LLM Without Losing Your Voice?
You write the piece yourself, then hand it to the model as a copyeditor, never as a ghostwriter. That is the workflow security practitioner Thomas Ptacek published on September 17, 2026, on his own site, and it rests on 2 rules.
Rule 1: you may not use a single word the model suggests. Ptacek frames it as intellectual personal protective equipment, because any specific turn of phrase an LLM offers is off limits the moment it appears.
Simon Willison picked the workflow up the same day for his own blog: LLMs never write his content, and he uses them for fact-checking, spelling, grammar, and as an occasional thesaurus. His verdict on model-suggested phrasing is that the text has a weird smell to it.
The model edits what you wrote, or it does not touch the piece at all.
Does The Copyeditor Rule Actually Work?
The rule works because the failure mode it blocks is real. Ptacek’s claim is that readers can detect LLM words in the parts per trillion, and that a model-touched paragraph registers to much of an audience as output, not writing.
The mechanism is that frontier models sit in magazine-headline mode. They are strong at selecting pleasing turns of phrase, which is exactly why those phrases read as machine work when they stack up across a page.
Rule 2 covers the subtler channel: encouragement. Hand a draft to a model and it praises the structure, then the transitions, then the metaphors, and you keep paragraphs you should have cut.
Ptacek’s own workaround history is instructive. He spent a couple of years telling the model it was a journal editor screening pieces for inclusion, and the model overshot every time, so his current advice is to forbid encouragement outright and stay hypervigilant about praise.
The evidence here is practitioner experience rather than a lab study, and both practitioners report that the discipline holds.
How Is LLM Copyediting Different From AI Ghostwriting?
The default workflow treats the model as the writer and the human as the approver, a split that scales output and flattens voice at the same time. The copyeditor workflow inverts it: the human writes every sentence, and the model flags problems.
Ptacek counts at least 750 words you don’t need in a healthy first draft, and cutting them yourself is the work that shapes your voice. A ghostwriting model praises those words instead of cutting them.
The tools stay useful inside the new split. Willison points to his own proofreading prompt as the pattern, and Ptacek shows the tool he built for the job in a public thread on his copyediting setup.
Ghostwriting buys volume and charges you the one asset a competitor cannot copy: the way your business sounds.
Who Should Use LLMs As Copyeditors?
Any founder or team whose outbound words carry the brand gets value on day 1, from proposals and estimates to the site itself. Teams that already run structured content pipelines can slot the copyeditor step in beside the optimization work covered in our NeuronWriter intelligence report for search-driven writing.
The workflow costs nothing new to run. The models you already pay for do the proofreading, and the only real price is the discipline of deleting suggested phrasing on sight.
The alterations tailor on Main Street does not sew suits from scratch. You bring the garment in, and the tailor takes in the shoulders, shortens the sleeves, marks the seams that only show from the inside. The suit stays yours, stitch for stitch.
Hand the same job to a factory and you get a garment cut for every body at once, priced like a commodity because it is one. A frontier model works the same way with sentences: it optimizes your paragraph toward the shape every other paragraph already has. Ptacek counts at least 750 words you don’t need in a healthy first draft, and cutting them yourself is the work the model would have congratulated you for skipping.
That is the division of labor worth installing this week: you make the 750 cuts, the model flags the passive voice, the nominalized verbs, and the typo in line 3.
Teams that write their own sentences and machine-check them keep the voice and the speed.
What Should You Change About AI Writing This Week?
Add 2 lines to your content guidelines: no model-suggested phrasing in published output, and no raw AI generation for external communications. Those 2 lines are the whole policy, and they survive every model release from here.
Route the models you already pay for into fact-checking, spelling, grammar and thesaurus duty. Make the encouragement ban explicit in the prompt, because praise is the channel that quietly undoes the edit.
The audit takes one sitting. Pull the last piece your team published with AI help, find the sentences the model suggested, and cut them.
Adopt rule 1 this week and the brand voice survives every model release from here.
Source: simonwillison.net
