Thomas Ptacek lays out two rules for writing with an LLM, use it as a copyeditor and never take a word it suggests

Thomas Ptacek posted a piece on 17 September 2026 arguing that LLMs help writing when used as a copyeditor and hurt it when used as a ghostwriter. The post reached 75 points on Hacker News in seven hours. Ptacek writes at sockpuppet.org and is a co-founder of Latacora and a well-known voice in the security community. The piece proposes two rules a writer should follow when running any draft through a model, and explains what a model can safely do once those rules are in place.

Rule one, never take a word the model suggests

Ptacek's first rule is that any specific turn of phrase the model produces is off limits, even when the reader thinks it is better than what they wrote. His argument is that frontier models are trained to select pleasing turns of phrase, and that a reader can detect model wording in the parts per trillion. Any paragraph written by the model reads as output rather than writing. Adopting the rule as a full ban is the safer move, because a writer will miss some of the ways a model would nudge a piece toward a house style the writer never chose.

Rule two, forbid encouragement

The second rule is to strip praise out of the interaction. Ptacek's point is that a model will tell a writer their draft is good when it is bad, and that this is worse than being unhelpful, because the writer then doubles down on first-draft impulses instead of doing the rethinks and rewrites that carry their voice. He writes that for a long time he opened every editing prompt with the lie that he was not the author but instead the editor of a publication screening a submission, to keep the model honest. He now recommends explicitly telling the model to skip encouragement and to be hypervigilant against any praise that slips through anyway.

What the model is genuinely good for

Once the two rules are in place, Ptacek argues the model is very good at flagging problems the writer cannot easily see by hand. Overuse of passive voice or nominalised verbs. Words like "very", "unfortunately", "really" and "actually" sprinkled through the draft. Paragraphs that would improve the piece if moved elsewhere. He recommends reading the book Style: Lessons In Clarity And Grace by Joseph Williams, taking notes from it, and turning those notes into a set of editing prompts the writer runs on every draft in order.

The suggested loop is to ask the model to spot problems, rewrite the flagged sentence or paragraph, then present the original and the rewrite to a second model with no context of the editing session and ask which is better. That second step exists because a model that knows the writer just rewrote something will tell them the new version is better regardless.

The reason a team building software would look at this is that the pattern is exactly the model-review pattern already common for code. A person writes, a model reads and flags, the person edits with the flags in mind. The value comes from the discipline in the middle, not from the model. Ptacek closes with a small demonstration of the rule: he fed the piece to GPT-5.6 with the lie that he did not write it, was told the essay was 20 percent too long, agreed it was probably correct, and did not shorten it.

Source

How To Write With An LLM by Thomas Ptacek at sockpuppet.org. Hacker News discussion at item 49747070.