Thomas Ptacek’s method fits in one sentence: use the model as a copyeditor, not a ghostwriter. Write the piece yourself, then feed it to a model to find flaws. Give the model the author’s chair instead, and readers register the result as output rather than writing — he claims “readers can detect LLM words in the parts per trillion,” however much you scuff the text into shape.

The two rules matter more than the workflow, because they are what keeps a useful tool from quietly rewriting your voice into something feeless and generic.

Rule one: you may not use a single word the model suggests

  • Frontier models are “supernaturally good at selecting pleasing turns of phrase.” That is the problem, not the benefit. Ptacek’s analogy: they are stuck in a mode where everything they write is a magazine headline, and you would wonder about someone whose article was made of dozens of headlines.
  • So any phrase the model offers is disqualified — even if you like it, even if you are sure it beats what you have. He frames this as “intellectual personal protective equipment”: you will not reliably spot which suggestions are the ones that make the writing worse, so take none of them.

Rule two: forbid encouragement

  • Hand an LLM a draft and it tells you the draft is great. That is the wrong feedback at the wrong moment. Your first draft has bad paragraphs, incoherent flow, and roughly 750 words you do not need.
  • The damage: praise at the structure stage makes you double down on first-draft impulses instead of doing the editing you would normally do. Ptacek calls those rethinks “load-bearing parts of your voice.” Readers cannot name what changed, but they sense the voice has been artificially flavored.
  • One procedural fix: when comparing your original paragraph to your rewrite, use a model that does not know you just rewrote it, otherwise it will prefer the new version to please you.

What these things are genuinely good at

  • Mechanical proofing, at volume: overused passive voice, verbs buried inside nominalizations, repeated tics (“very,” “really,” “actually”), and the two or three paragraphs that would be clearer moved elsewhere. You can find all of this yourself — it is just tedious, and the model does not get tired.
  • His recommendation for a schematic of those edits is Style: Lessons in Clarity and Grace, which he learned about from Richard Gabriel and describes as turning copyediting into programming. Read it, take notes, turn the notes into prompts, run them as passes.
  • He built a small workshopping tool to run those passes — a prose editor with highlighting and Genius-style sidebar commentary, stepping forward and back through suggestions — and drives it through the Codex, Claude, or Antigravity command-line tools.
  • The essay refuses its own advice at the end. GPT-5 reviewed the piece and said it was 20% too long. “It’s probably right. But I’m not fixing it. I’m just gonna be me.”

What the thread adds

  • tptacek — replying as the post’s author with the system prompt behind the method, after someone asked how his tool actually prevents the model from praising him: “Workshop a piece with me. NO ENCOURAGEMENT. Encouragement is useless; the only useful things are suggested corrections. DO NOT WRITE COPY FOR ME. Any words you provide will be disqualified, so if you come up with good words, I’m fucked because I can’t use them.” He also supplies the calibration the essay leaves out: “Assume I do 8-12 passes on any piece. Assume I respond to ~60% of suggestions; the other 40% I’m like, ’nah, I wanted it to sound that way’.”
  • simonw — the one use he calls a clear win is fact-checking rather than style: “It’s shocking how well it works… today the good ones (Fable, GPT-5.6/6) with a search tool enabled are genuinely excellent at this. They can even retrieve articles you link to and check that you aren’t misrepresenting them.” thombles raised the prompt first, and reports one case where it pointed him at documentation that undermined his entire post.
  • semiquaver — extends the rule into engineering writing: they now insist on writing all their own commit messages and pull request descriptions, letting an agent check those for factual accuracy but never rephrase them. “It has helped immensely to improve the depth of my understanding of the agent-generated code. When agents are doing everything its way too easy to ‘skim’ diffs and not really absorb them.”
  • awithrow — how to tell when the model has run out of real advice: “it’ll offer the same ’trim 10-15%’.” Their test is to submit other people’s writing; run it enough times and you see the same generic edits applied to almost any input.
  • Retr0id — the dissenting technical objection: “LLMs have absolutely terrible taste when it comes to writing. I don’t find their feedback useful at all, beyond trivial spelling/grammar mistakes… Proofreading is all you need.” Ptacek’s reply restates the whole essay in one line: “That’s what this is: it’s mechanical proofing.”
  • mjmvisser, vidarh, and thunfischtoast — three separate commenters stopped on the same phrase. Quoting “those rethinks are load-bearing parts of your voice,” mjmvisser posted rule one followed by “So far so good… Oops.” vidarh: “I really hope this was intentional.” It was — Ptacek confirmed “It sure was” — which makes it a nice demonstration of the article’s own point about how quickly readers start scanning human prose for machine fingerprints.

Where the thread pushes back

  • Several commenters argue the honest advice is simpler than two rules: do not use a model for anything a human will read. Eddy_Viscosity2 puts the line at the audience — “if you’re writing for a human mind to ingest and extract meaning from, then LLMs are poison” — while ttctciyf goes further, calling AI “a stupidity amplifier” in general use.
  • S-E-P describes the reader-side cost of the practice: “I still, while reading, am checking for any amount of AI used. This makes reading less enjoyable, and even a bit stressful. If you can’t spend the time to write it, why should anyone read it?” jakeinspace and iqp agree in the replies, with iqp adding that more content is being produced faster than ever and most of it “won’t stand the test of time.”
  • bsenftner offers the conditional version of the thesis: write your own revision from the critique, and slop is impossible “unless your own writing is already in the style of LLM output.”
  • brlewis agrees with the central thesis but not all the details, and wrote a blog post response to say where.

On reading comments as evidence: HN handles are pseudonymous, the site publishes no per-comment scores, and the order here is HN’s own ranking — so this is a slice of the thread, not a consensus.