Send Me Your Prompt

I should start with a confession, because this essay would be dishonest without it. Most of what you’re reading was drafted by an AI agent. It searched the posts I’d saved on X, read the essays they pointed to, and turned my notes into paragraphs. The map this comes from, What We Had to Unlearn, was made the same way, and its source note says so.

That matters more here than anywhere else. The map’s chapter on writing ended by saying readers want to know that a person meant every sentence. By the standard of almost everyone quoted below, a draft like this one is the problem.¹/01field noteCantrill’s test is whether readers can tell, Clay’s is whether you stand behind every sentence, and Ptacek’s is whether a single word came from the model. An agent-drafted essay fails the first and third until it’s rewritten.Thomas H. Ptacek, Sep 16 · Bryan Cantrill, “The revolt of the reader” · Clay’s AI writing policySo I’m going to use it as the test case. If the rule I end up with can’t say what’s wrong with this draft, and what would fix it, it isn’t a rule.

Here is the rule, up front. Compress, don’t expand. AI that shrinks your text (comments on a draft, an edit, a summary, a translation of something you already thought) keeps your thinking. AI that inflates a prompt into prose is the offense. Everything else in this essay is the argument for that line, and the cases where it gets hard.

The writer paid more

The odd thing about this year is that two rules moved in opposite directions. Engineers who stopped reading every line of code drew a harder line than ever around prose. Pamela Fox put both halves in two sentences. One reply explained the difference better than anything I’ve read since: code has tests to hide behind, and prose has nowhere to hide.²/02the postsone paragraph of laminated corporate oatmeal and the whole README is guilty. ND Minds & AI, Sep 4checking a draft takes longer than writing one. ▪️Wayne▪️, Sep 6Please do a human pass on READMEs and docs, for all of our sakes. Pamela Fox, Sep 4

The rule underneath is old, and nobody used to need to say it. When you read something, someone wrote it, and they worked harder on it than you will. Bryan Cantrill calls this a social contract in RFD 576, Oxide’s guide to using LLMs: “it is more work to write than to read!” If you struggle with an idea, you can assume the writer understood it, so the least you can do is labor to make sense of it. Generated prose rips that up. The reader can’t assume the writer understood anything, “because they might not so much have read the product of the LLM that they tasked to write it.”³/03field noteRFD 576 is stricter about personal writing than public writing: for chat, review comments, and e-mail, models may inform or edit, but “should never be used to write text that the reader is very clearly expecting to be human-authored.”RFD 576, Using LLMs at Oxide

Clay’s AI writing policy, written by Sophie Alpert, turns the same idea into arithmetic. One person writes a document and many people read it, so every minute a reader spends untangling it gets multiplied. “More time should be spent authoring a document than consuming it.” Wayne de Jager put it more bluntly in the replies to Gergely Orosz: what people resent isn’t the prose, it’s that reading it cost more than sending it did.

Readers started acting on it. In The revolt of the reader, Cantrill cites a survey of 668 developers by Cynthia Dunlop: 78% said they stop reading as soon as they detect an LLM, 71% avoid the author afterward, and 98% would rather read an author’s own imperfect piece than an LLM-polished one.⁴/04field noteThese figures come through Cantrill’s essay; I haven’t read the survey itself. Cantrill anticipates the self-selection objection and argues that active readers on social media are exactly the people who promote writing they like.Bryan Cantrill, “The revolt of the reader” Those are big numbers from a self-selected group. I haven’t read the survey itself. But I recognize the reflex, and so did a lot of people in the replies.

What broke is a signal. Fluent writing used to be expensive. It took hours to squeeze a three-dimensional idea into a line of words someone else could follow, so a fluent piece was decent evidence that someone had done the thinking. Juliette, in the same week, put it as plainly as anyone: fluency is no longer a measure of thought.⁵/05field note“A fluent piece of writing can still mean that many hours of thought went into it. It can also mean a one-line prompt did.”Juliette, Sep 6

the broken signal

Fluent no longer means you thought about it.

Send me the prompt

The map noted something I still find striking. Over about two months, at least four people landed on the same request without, as far as I can tell, coordinating. Kenton Varda in July, about pull request and commit descriptions: “I’d rather see your prompt than your output.” Clay’s policy in August: “consider instead just sharing the prompt itself.” Cantrill in September: “just give us your prompt.” And Peter Steinberger, from the receiving end of a pull request, asking why the prompt had to go through two agents at all.⁶/06field noteThe four, in order: Varda on July 8, Clay’s policy on August 19, Cantrill on September 5, and Steinberger on September 7.Kenton Varda, Jul 8 · Bryan Cantrill, Sep 5 · Peter Steinberger, Sep 7

What the map didn’t do was say why the prompt. Here’s my answer. The prompt is the compressed thing. It’s the part that came from a person: what they wanted, what they thought mattered, in the order they thought of it. Everything the model adds on top is the prompt plus padding. If the reader wanted padding, they have a model too.⁷/07the postsmost of the text is bloat: either repetition or LLM world knowledge Felix Huttmann, Sep 5

Felix Huttmann named it best, in a reply to Cantrill with one like. LLM text is fine if it has something interesting to tell. The problem is when a prompt is expanded. Then most of what you read is either repetition or general knowledge the reader could have gotten from their own chatbot. Clay’s policy says nearly the same thing: if you generate a document from a short prompt, “They can always talk to ChatGPT themselves if they want to.”

Read Varda’s original post again with that in mind. The complaint wasn’t that the descriptions were wrong. It was that they went in the wrong direction. They described “details of the code that could easily be seen by looking at the code,” and left out “the higher-level framing needed to understand broadly what the code is doing.” The model expanded the part the reviewer already had and dropped the part only the author knew.⁸/08field noteVarda’s diagnosis of why people like the output anyway: it “looks structured and thorough, which makes it feel professional.” Nuxt’s contribution guidelines put the same rule in five words: never let an LLM speak for you.Kenton Varda, Jul 8 · Alexander Lichter, Jul 8 That’s what expansion does. It’s very good at adding what everyone knows and very bad at keeping what only you know.

None of this is new. Clay’s policy quotes Pascal: “I have made this [letter] longer than usual because I have not had time to make it shorter.” Writers have always known that length is the cheap part. The model just made the cheap part free.

the rule, said rudely

If you had an AI write it, send me the prompt. The rest is padding.

Editors, not writers

If expansion is the offense, what’s allowed? The answer most of these people gave is the same one: let the model edit, not write.

The clearest version came out of an argument. In late August, Unazi Ogah challenged Cantrill about Oxide’s use of Pangram, an AI-text detector. Write a complete text yourself, Unazi said, then ask a model to fix the grammar and paragraphing, and Pangram will likely call it 100% AI-generated. Do you want a tool that denies people authorship for fixing their grammar? Cantrill’s answer drew the line in one sentence: use the model for comments, not rewrites.⁹/09field noteWhether a grammar fix alone gets flagged is an open empirical question: Unazi says it will, Cantrill says comments-only stays human. Unazi also argued that no false-positive rate is acceptable, given the cost of a false accusation.Unazi Ogah, Aug 29 · Bryan Cantrill, Aug 29¹⁰/10the postsA false-positive result, no matter how small, is not acceptable Unazi Ogah, Aug 29LLMs are phenomenal editors … you will find that your writing remains human-authored. Bryan Cantrill, Aug 29the single nerdiest writing book ever written Thomas H. Ptacek, Sep 16If I can’t defend every sentence from work I actually did, it doesn’t get posted. Alain Manders, Sep 7

RFD 576 says the same thing at more length. LLMs can be excellent editors, late in the process, once a document is “already written and broadly polished.” The earlier you bring them in, the more likely they are to steer you somewhere else, “at once praising your groundbreaking genius while offering to rewrite it for you.”¹¹/11field noteA designer’s version of the same rule: no AI for the first draft, because the first draft is where the deep thinking happens; feedback from a model on the second or third is fine.Shivam Dewan, Sep 21 · RFD 576, “LLMs as editors” Thomas Ptacek’s version is the strictest: use the model all you want, but don’t keep a single word it chose. His guide comes from Joseph Williams’s Style, a book Ptacek calls one made for a moment when robots can be our copyeditors.¹²/12field notePtacek’s reason for the zero-words rule: readers “detect LLM words in the parts per trillion.” It isn’t anti-AI: “There’s no reason you can’t use an LLM to help you write.”Thomas H. Ptacek, Sep 16 Alain Manders draws the line by what the model may touch: it may organize and challenge the thinking, but it may not invent the experience.

Lay those out and they sort by direction, not by whether a model touched the text.

The best evidence for this line came from an experiment that wasn’t about it. Paul Novosad ran old French and Spanish writing through a series of AI translation prompts and checked each result with Pangram. Translation alone didn’t get flagged; the human voice survived the change of language. The only way Novosad could get a “100% AI” score was to cut a Piketty op-ed down to bullets and then have the model write it back up from the bullets.¹³/13the poststhis of course eliminates Piketty's voice completely. Paul Novosad, Oct 2

Compression, then expansion. That’s the thing the detector saw, and it’s the thing readers see too.¹⁴/14field noteNovosad’s conclusion from the experiment: “distinct human voices are distinct even after language translation.” Whether we care about a human voice depends on the context.Paul Novosad, Oct 2

I should concede what Clay’s policy concedes. “There are no lossless transformations of natural-language text.” Even a shrinking edit changes the meaning a little, and the model doing it doesn’t know what you meant as well as you do. So this is a rule about degree. A copyedit loses almost nothing. A summary loses some. A prompt grown into an essay loses the writer.

the line

Compression is fine. Expansion is the crime.

The case for prompting better

The other side deserves its strongest form, and it has some of the most-liked posts in this whole argument.

David Cramer replied to Varda’s moratorium with the exact opposite: never writing a commit or PR description again, and building a skill to make the output repeatable. Another reply, one Varda answered, made the bolder claim: AI can follow instructions better than a human.¹⁵/15field noteThe reply Varda answered asked why not just give the model a rubric or skill for descriptions: “AI can instruction follow better than human.”George, Jul 8 Varda’s rebuttal is the one I keep. The description isn’t paperwork. It’s evidence that a human understood the change. “If human developers can’t even write two paragraphs describing the code they are submitting then we have a bigger problem.”¹⁶/16the postsim the entire opposite: never writing a commit/PR description again David Cramer, Jul 8Even if AI is writing the code, we (my team at least) still need humans to understand what is being built. Kenton Varda, Jul 8

Notice Varda’s concession, though. For pull requests made fully autonomously, Varda says, yes, prompt the AI to write better descriptions. That isn’t a contradiction. When no person did the thinking, there’s no person’s thinking to lose.

The most popular post on this side, with nearly 27,000 likes when I captured it, said you can get good technical documentation out of a model by making it follow ASD-STE100, Simplified Technical English. Sujayath’s reply is the best answer I’ve seen. STE was built for aircraft maintenance manuals, read by technicians who often aren’t native English speakers. It’s voiceless by design. The complaint about AI writing isn’t that it’s too colorful. It’s that it’s generic. “You would be fixing AI slop with a different flavor of AI slop.”¹⁷/17the postsFlat, hedged, saying nothing, sounding like everyone else. STE makes writing more uniform, not less. Sujayath, Aug 9what if the text is LLM-written with substantial human effort? Marco Giancotti, Sep 6

The fairest objection came from Marco Giancotti, who said nobody had ever answered it. Suppose someone puts in 90% of the effort of writing it themselves and gets 200% of the quality, because they know they’re a poor writer. Is that still worthless? Where is the line?¹⁸/18field noteMarco Giancotti’s fuller case: the average person writes worse than a modern model, and someone can prompt, iterate, cut, and add by hand. “It feels like throwing the baby out with the bathwater.”Marco Giancotti, Sep 6

Here’s my answer. The line isn’t effort, because the reader can’t see effort. The line is direction. If Marco’s 200% is the writer’s own thinking, stated more clearly than they could manage alone, with the model cutting and tightening, that’s compression, and it’s fine. If the 200% is the model’s ideas and structure, with the writer approving them, that’s expansion, however long it took. The honest test is Clay’s: could you answer “What did you mean by this line?” for every line, without “Oh sorry, AI wrote that”?

Prompting better fixes the style. It can’t put back the thinking. That’s also why I don’t expect better models to settle this, though plenty of people expect them to.¹⁹/19field noteClaire Vo listed “writing like a human” among what a new model does significantly better. Clay’s policy hedges the same way: as models improve, “it may make sense to lean more heavily on them, but the principles in this policy will remain important.”claire vo 🖤, Jul 8 · Clay’s AI writing policy A model that writes exactly like a human makes expansion harder to catch. It doesn’t make it something other than expansion.

Who gets to write it themselves

There’s a fairness problem in all of this, and RFD 576 is unusually honest about it. Oxide’s hiring process “very much selects for writers,” so everyone there can write, and “we have the luxury of demanding of ourselves the kind of writing that we know that we are all capable of.” That’s a rule written by people who can write well, for people who can write well.

Plenty of people can’t, or can’t in English. Some of them pushed back hard.²⁰/20the postsarticulating thoughts in words isn’t a skill everyone possesses. LLMs helped this immensely and I am grateful. ☷ = ☰, Sep 6A detector can inspect words, but it cannot inspect authorship. Antonio Grasso, Sep 14I'd rather read grammatically incorrect human text than fluent AI text any day. Talia Ringer 🕊🪬, Sep 29

Antonio Grasso’s post is the one I’d hold every rule up against. Grasso thinks in Italian, works out the idea in Italian, then explains it to a model, challenges its words, and rejects whatever doesn’t sound like Grasso, until the English says what was meant. Then a detector calls it 100% AI-written. “Ma fatemi il piacere,” roughly: oh, give me a break.

So does the rule protect Grasso? It should, and I want to say so plainly. Grasso’s process is compression of Grasso’s own thinking into another language, the same move Novosad’s experiment showed a good detector lets through. The same goes for someone with dyslexia who knows exactly what they want to say and uses a model to get it onto the page. If the ideas, the structure, and the emphasis are yours, and the model shrinks, fixes, or translates them, you wrote it. Paul Novosad expects non-native speakers to lean on AI more and more for their English, because the cost advantage is so large.²¹/21field noteNovosad, a few weeks after the experiment: “it seems inevitable that non-native English speakers will increasingly rely on AI to do their English writing.”Paul Novosad, Oct 22 A rule that can’t tell that apart from a prompt grown into an essay is a bad rule.

The rule still asks something of them, though. It asks the same thing it asks of everyone: don’t let the model add ideas you didn’t have. Talia Ringer’s reply is the uncomfortable half. Fluent isn’t the goal. Ringer would rather read grammatically incorrect human text than fluent AI text. I think most readers would, too, if they knew which one they were looking at. That’s the hard part.

The detector as judge

Cantrill’s answer to “which one am I looking at” is a detector. In the essay, Cantrill compares LLM prose to e-mail spam. In the early 2000s spam looked like it might kill e-mail. The war turned once filters could identify spam at scale, and being labeled spam became a disaster for a sender. Cantrill found Pangram’s newest model good enough that Oxide now requires its public writing to come back from Pangram as human-authored. “You should fully expect your writing to be run through Pangram.”

The obvious objection is false positives, and the numbers there mostly come from the vendor. Max Spero, Pangram’s co-founder and CEO, has claimed a false-positive rate of 1 in 10,000.²²/22field noteTreat that number as a vendor claim. RFD 576 cites the Pangram 4 technical report (arXiv:2607.27183) for its own confidence; I haven’t read the report, or checked whether it reports error rates for non-native writers or edited text.RFD 576, “LLMs in public writing” Others found the tool brittle at the edges: a short abstract the author translated without AI, flagged anyway;²³/23field noteA reply to Novosad’s thread: abstracts translated by hand still came back flagged. “Perhaps long texts are not, but beware short texts like abstracts.”Manoel Galdino, Oct 3 a human essay that turned 25% AI with a one-word section title.²⁴/24the poststhey're brittle and false accusations can cause immense harm. Isaac Flath, Aug 21a committee decides that Jane is not worthy of a job, because Pangram told them to ignore her. Yonathan Arbel, Oct 8You can't even say "it's not just x, it's y" anymore without it smelling like an LLM. Gregory Wieber, Sep 6Back then, their writing would just be called poor English. Now it’ll be called AI-generated. T. Y. Hayashi, Sep 18Max Spero, Mar 22 · Paul Novosad, Oct 2

Isaac Flath’s complaint is fair to Pangram, which Flath calls “a fantastic classifier.” The problem is what happens downstream. A teacher sees a score, has no way to reason about it, and the student has no one to appeal to. T. Y. Hayashi said it in two sentences: non-native scholars used to be told their English was poor; now they’ll be told it’s AI-generated. Yonathan Arbel, writing about hiring committees in law, made the point that matters most for my rule. Pangram gives the committee a score. It can’t tell a paper generated from whole cloth from one that got help fixing run-on sentences.²⁵/25field noteSecondhand, via a summary I haven’t traced to the source: an MIT report reportedly recommends against relying on AI detectors, citing false positives on non-native and neurodivergent writers. One reply to Cantrill: “I’d trust an edit history more than an AI detector.”Rohan Paul, Aug 27 · Andrea Postiglione, Sep 7

That’s the gap. A detector measures the text. My rule is about the direction the text moved. A good detector is a proxy for that, and Novosad’s experiment suggests Pangram is a decent one. But a proxy can’t be the judge, and Gregory Wieber noticed the side effect: people now write around the tells, avoiding perfectly human phrases just so nobody dismisses them. Cantrill, a self-described “unapologetic” user of the em-dash, has the same worry about cruder detectors.

There’s a tension inside RFD 576 itself that I don’t want to paper over. The same document has a section on “LLM shaming,” which calls it “decidedly anti-social” to out other people’s LLM use, next to a rule that makes a detector score a condition of publishing.²⁶/26field noteRFD 576 compares LLM use to a dietary choice: accommodate vegetarians, but “it is decidedly anti-social to interrupt someone’s meal to register disapproval with their choice of entrée.”RFD 576, “LLM shaming” I don’t think that’s hypocrisy. One rule is about how you treat colleagues, and the other is about what Oxide signs its name to. But it shows how hard it is to police the text without policing the person.

Show, don’t tell

Here’s where I’ve landed, and it’s the part of this essay I’m most sure of.

AI writes excellent prose for other AIs, and dog shit for people. A human always expresses themselves more legibly than a model does. Always. That isn’t a complaint about grammar. Model prose is fluent. It’s a complaint about who it’s for. A model writes the way a model reads: everything stated, nothing assumed, every point repeated in case it was missed. That’s great when the reader is another agent. It’s exhausting when the reader is a person who wanted to know what you think.

two audiences

AI prose is great for machines and dog shit for people.

So split the writing by audience. RFD 576 already has a category for this, operational writing, the documentation and RFP answers where “the reader’s expectation is correctness, not authorship.” Clay’s policy allows quoting AI output verbatim as long as you mark it. And gfodor, in the replies to Cantrill, started labeling a second kind of essay: written largely by LLMs, for agents to read. Mike Manzano’s reply to Pamela Fox suggests more of the world is moving into that category than we admit: READMEs are increasingly read only by LLMs.²⁷/27field notegfodor’s label read: “Feel free to point your agent at it instead of reading it yourself if you want to digest it.” Clay’s policy allows the same move inside a document: “Claude offered this idea, do you think it’s worth looking more into?”gfodor.id, Sep 5 · Clay’s AI writing policy²⁸/28the poststhe real bug is provenance … your name stops working as a checksum Jayant Lohia, Sep 7Mike Manzano ✨ ᯅ 🔳, Sep 5

Labeling is what keeps Jayant Lohia’s checksum working. Your name on a piece of writing is supposed to say which sentences you checked. Once readers can’t tell, the name stops meaning anything, on the labeled pieces and the unlabeled ones alike.

But agents still have to report to people. A pull request, a fix, a status update: somebody human has to understand it. That’s where the rule turns into advice. When an agent has to reach a person, it shouldn’t write more. It should show. A diagram, a screenshot, a recording of the thing working.²⁹/29the postshow differently do you approach a code review if you know there's video proof that it works? Thorsten Ball, Oct 13verify work end-to-end with screenshots and recordings Ben Holmes, Nov 11

The people building with agents every day got there on their own. Thorsten Ball had an agent record a video walkthrough of its own change and asked how differently you’d review a change with video proof that it works. Ben Holmes has agents verify their work end-to-end with screenshots and recordings before opening a pull request. A summary of a talk by poteto, posted in Chinese, lists the same rule among the habits of a team shipping thousands of pull requests a month: a fix must come with screenshots or video, not a little essay claiming it works.³⁰/30field noteThe summary, in my translation: fixes must come with screenshots, video, or other evidence; don’t rely on a little essay to prove you got it right. Others list “proof that the change works on every PR (video/screenshots)” as a guardrail, and Thorsten Ball notes models can now generate explainer videos and walkthroughs for any code.高运通, Nov 12 · Oskar, Nov 9 · Thorsten Ball, Oct 16 The pull requests for this site work the same way: a rendered diagram up top for me, and a collapsed section of text for whatever agent reads it next.

A picture can’t pad. A screenshot either shows the bug fixed or it doesn’t. A diagram either matches the system or someone notices it doesn’t. Video is hard to fake into seeming thoughtful. That’s why visuals are the right way for an agent to talk to a human: they’re compression by nature.

And when a person has to talk to a person, write it yourself. Rob Nelson said why better than I can: if I wanted to ask an LLM, I could have. If I’m reading your writing, I want to know what you think.³¹/31field note“Meat proxy” also took on a second meaning this year, for review comments relayed between agents and for jobs reduced to that relay. That sense lives in the map essay’s writing chapter.What We Had to Unlearn, “Writing it yourself”³²/32the postsDon't be a meat proxy! Rob Nelson, Sep 6

Which brings me back to this draft. By the rule I’ve just argued for, what you’re reading is expansion: my notes, a cache of posts, and a few paragraphs of my own words, grown by an agent into an essay. It has the virtue of being labeled, and the vice of being exactly what the people in it complain about. What would fix it is the thing every one of them asked for. Either I send you the notes, or I do the work of making it shorter, in my own words, until every sentence is one I’d defend if you asked what I meant by it.