How to Write with an LLM
TL;DR
Write every word yourself, and use an LLM only as a copyeditor that flags problems. Two rules keep it from flattening your voice. First, never accept a single word it suggests. Second, forbid it from praising your draft. Its encouragement talks you out of the rewrites good writing needs.
The article
- Readers spot LLM prose at tiny concentrations and file it as output rather than writing. So tptacek says the drafting has to be yours. The model's job starts after the draft exists.
- Rule one bans every word and phrase the model suggests, even ones you like. Frontier models write like magazine headlines, and you won't reliably catch the ways that drift into your text.
- Rule two bans encouragement. A first draft is mostly bad. A model that praises its structure, transitions and metaphors pushes you to keep choices you would otherwise rewrite.
- Models are good at the tedious mechanical checks. These include passive voice, verbs buried in nouns, repeated phrasing, filler words like "very" and "actually", and paragraphs that belong elsewhere. Joseph Williams's book Style gives a checklist you can turn into prompts and run as separate passes.
- To compare an original paragraph with your rewrite, ask a model that hasn't seen your editing. One that has will guess you want the new version to win. tptacek built a small workshop app that runs each editing prompt through the Codex, Claude or Antigravity CLIs.
Why it matters
Worth reading for anyone who drafts essays or technical posts and wants faster editing without sounding machine-made. The original has the exact app-building prompt. In the thread, tptacek posted his system prompt, with lines like "NO ENCOURAGEMENT" and "DO NOT WRITE COPY FOR ME". He said he runs 8 to 12 passes and acts on about 60% of the flags.
The discussion
A large and often heated thread, where many commenters skipped the article's advice and argued whether anyone should write with an LLM at all. A long run of one-word "Don't" replies drew the reply that the piece says the same thing. The author, tptacek, was in the thread repeatedly, insisting he asks models for technical flags, not style opinions. He confirmed that his use of "load-bearing" was a joke.
Why read what nobody bothered to write
The biggest thread opened with a reader who now scans everything for AI, asking why anyone should read text its writer didn't spend time writing. Defenders said the value is in the ideas and that politicians and professors have long had others polish their words. Critics answered that polishing strips sincerity and that LLMs judge badly what a reader already knows. A side fight over the aphorism's own originality ran for dozens of comments, and several described AI-written Jira tickets and docs they stopped reading.
If you can't spend the time to write it, why should anyone read it?
Because the value of writing are not the words on a piece of paper but the idea they convey.
Commit messages and PR descriptions
The top commenter went back to writing his own commit messages and PRs, and found it made him actually absorb agent-written diffs. Some replies said LLM messages beat the old one-word "fix", while others called them bloated jargon that still sends you to the diff. Several agreed the useful part is the why, which a diff summary can't supply. One team's only AI policy, no AI PR descriptions, got praise for forcing authors to understand their own changes.
It establishes personal ownership for the submitter, making it psychologically difficult to submit something you don't understand.
I think that the idea of generating a commit message based on the content (diff) of the commit is fundamentally wrong.
Can a model judge prose at all
Skeptics said LLMs pick LLM-style writing in A/B tests, so asking which version is better steers you toward their prose, even if you wrote both. Others replied that model style isn't the opposite of quality and that a careful reader with different blind spots still helps. tptacek kept repeating that his prompts flag checkable things like overused words, throat-clearing and buried verbs, and that he decides what to change. Some pointed to existing tools like Hemingway, LanguageTool or iA Writer's style check, and he asked which of them finds nominalizations or weak topic sentences.
By A/B testing against an LLMs opinion, you are optimizing in the direction of LLM prose even if the LLM never writes any of the prose itself.
More importantly: they're technical. They're not really judgement calls.
The "load-bearing" joke
At least a dozen threads pounced on "Those rethinks are load-bearing parts of your voice" as an AI tell in a post banning AI words. Most guessed it was deliberate, and tptacek confirmed it. Some still said it pulled them out of the piece, and others mourned a useful word that LLM overuse has spoiled. One critic argued that even reading model feedback leaks its style into your head.
But while you're reading their output, the slop makes its way directly into your brain anyway - and then it drips out into your writing.
It was a joke.
Fact-checking and other narrow uses
Asking a model to fact-check a post drew the least pushback, and simonw said current models with search do it well. One commenter warned that a model was confidently wrong about their own code, and the answer was to check what it flags, not trust it. Others outlined voice-memo rambles into plain facts with a model, or used it as a thesaurus for one elusive word. Some drafted in another language and rewrote by hand, or asked how a hot take might be misread.
on at least one occasion it pointed me towards some documentation that undermined my entire post - and led to a much better one a week later once I'd processed and applied that new information.
LLMs are super handy as fancy vibe-based dictionary search engines.
Consensus
Unedited LLM prose is easy to spot and tiresome to read. If you write for people, the words and the thinking behind them should be yours. A model is at most a second pair of eyes.
Dissent
A sizable group rejected even the editor role. They said model feedback has poor taste and pulls you toward its style. They added that a friend or a night's rest does the job better. And writing is the thinking, so outsourcing any of it costs you.