Distracted by the LLM-generated style of writing. Not sure whether the author truly writes like that (unlikely) or there was heavy AI assistance in drafting this.
I have no idea about the current article, and given that the author is the person with the first commit in the Kubernetes repo (https://joe.dev/about/), he obviously has a lot of credibility.
Just generally though: what we're seeing a ton of these days is people writing something and then passing it to an LLM with a request to improve it somehow, e.g. by fixing grammar, tightening the style, etc. In such cases, the answer to your question is that the "prompt" is (1) a first draft, and (2) an instruction to edit it.
It's clear, though, that the LLMs leave far more imprints on the text than most people realize, and that although they may have asked the LLM to restrict its edits to "just" X or Y, the actual changes to the text will often go beyond that.
How this will evolve over time is anyone's guess, of course.
I have a steering .md file that instructs Opus how not to sound like an LLM when writing prose (I write prose in my IDE with Opus). The steering is specific to me, but I've found that giving Opus rules like eschewing punchy journalistic sentences ("Not because X. But because Y. And that matters."), varying sentence lengths and avoiding staccato sounding clauses go a long way in smoothing out LLM smells in writing (at least according to me).
Aside: different LLMs sound different too! ChatGPT is the worst offender for LLM-sounding writing and needs the most smoothing, but Claude (web) actually sounds like a humanities major from the get-go.
I do this so many times. Type in a large amount of text and the only thing final in my mind is para breaks and the idea per paragraph. And then give it to AI saying "sending to director", "sending to friends on WhatsApp group", "sending to colleagues" and it does an awesome job of bringing the "AI polish" and then you edit or negotiate line by line or para by para on what you want to keep.
Yes, this is certainly common, and opinions and tastes differ about the outcomes—as they should, because we are all still in the early stage of sorting through the best way to use these tools. I think it's also already clear that "best way" means something different in different contexts.
Where some people are getting into trouble, at least in the HN context, is underestimating the impact that this has on their text. There's a big perception gap between the author's view ("fixed up the grammar a bit") and the reader's view ("this sounds entirely like an AI wrote it") in many cases. So many, in fact, that I feel I can say something about it. I'm no authority on any of this and don't want to sound like one, but this is such a common pattern at the moment that I feel confident reporting it. How it will change over time, I have no idea.
(I also don't want to sound anti-LLM - we rely on these tools heavily, they're amazing, they've already improved HN, and they show every sign of high potential to improve it further. The bottleneck isn't the LLMs, it's how quickly we can figure out how to use (and test) them. We just don't use them to process any text that we put on HN itself.)
Attached is the writing.md I use to steer Opus 4.8. Prompt:
"use the writing.md steering on x.md and loop until all LLM traces are removed".
I ran it on TFA and Pangram flagged it as LLM generated but Claude Fable couldn't definitively tell.
-- writing.md ---
# Writing Rules (MANDATORY)
## Banned Words and Phrases
Never use: "incredibly", "extremely", "absolutely", "fundamentally", "dramatically", "crucial", "vital", "powerful", "robust", "elegant", "seamless", "cutting-edge", "game-changing", "groundbreaking", "It's worth noting", "Importantly,", "Interestingly,", "Let's dive in", "At its core,", "At the end of the day,".
No exclamation marks in technical writing. No contractions in formal writing.
## Banned Sentence Structures
1. *Semicolons joining independent clauses.* Do not write "X does A; Y does B." Use a comma + conjunction that names the relationship: ", while" (contrast), ", and" (addition), ", so" (consequence). Semicolons hide the logical link and sound artificially balanced.
2. *Label-colon-explanation.* Do not write "The key insight: ..." or "The limitation: ...". State the point directly or use "is that" phrasing.
3. *Colon after a bolded term.* Do not write "a *rollout engine*: a lightweight...". Use a comma.
4. *Sentence fragments as assertions.* Every claim needs a subject and verb. "No gap at any ρ." → "There is no gap at any ρ."
5. *Em-dashes joining independent clauses.* Do not write "X does A — Y does B." Use a comma + conjunction. Parenthetical em-dashes ("the policy — trained offline — cannot adapt") are fine.
6. *Tricolon lists of near-synonyms.* "It does not X, Y, or Z" is padding unless each item is genuinely distinct.
## Banned Rhythms
1. *Staccato sequences.* Two or more consecutive short declarative sentences of similar length. Join them with a conjunction or subordinate one. A single short sentence standing alone for emphasis is fine and often good. Do not eliminate it.
2. *Formulaic layout.* Do not produce: intro paragraph → three bullets → summary paragraph.
3. *Gratuitous parallelism.* Do not force list items into identical grammatical form if it makes them sound robotic.
4. *Saying it twice.* If you stated a fact, do not rephrase it from another angle in the same paragraph. One clear statement is enough.
5. *The negation-correction reversal.* This is the move where you deny one candidate and assert the real one. Surface forms to match: "not X, but Y"; "it isn't X, it's Y"; "X was never the point, Y was"; "for me X, for them Y"; the comma-tag "Y, not X" ("sanctioned, not stolen"); the "not so much X as Y" form; and the gapped version where a stranded verb delivers the pivot ("Hours aren't the bottleneck. Attention is."). One reversal at a genuine turning point is good writing. The structure is not the problem. The density is.
Detection is a whole-document pass, not a per-paragraph glance. Read the entire piece and mark every sentence or sentence pair that negates one thing to elevate another, including the comma-tag and gapped variants above. Count the marks. More than one per ~300 words, or more than three in a short piece, means the reversal has become the default sentence engine, which is the machine tell. A single dense paragraph with two stacked reversals also counts.
Fix by thinning, not by deleting all of them. Keep the two or three that land on the strongest turns. Rewrite the rest as plain declaratives that state the point with no foil ("The bottleneck is attention now." instead of "Hours aren't the bottleneck. Attention is."). Removing every instance flattens the voice, so the aim is to make the survivors rare enough to regain their force.
6. *Repeated hedge adverbs.* A softener like "almost", "somewhat", "rather", "a bit", or "fairly" used more than once in close range becomes a tic. Keep at most one, and only where it earns its place.
## Positive Rules
- Active voice. Use "we" and "our".
- Concrete nouns and verbs. "The model overfits after 50 epochs" not "exhibits suboptimal generalization characteristics."
- Plain English. Use technical terms only when they carry meaning plain English cannot.
- State consequences, not meta-commentary. "The policy has no lookahead" not "training compresses multi-period consequences into a single-step mapping."
- One sentence that advances to the next thought beats two sentences restating the current thought.
- State assumptions when uncertain. Do not hedge-stack ("it might be the case that perhaps...").
- Not every paragraph needs a topic sentence or a concluding sentence.
- Do not resolve the ending with a tidy bow. A piece may close on an open question, an admission, or an unresolved tension. Summary endings that restate the thesis read as machine-generated.
- Do not over-smooth. Removing every short sentence, every parallel, and every fragment flattens prose into uniform medium-length sentences, which is itself an LLM smell. Fixing a tell should not cost the voice.
- When editing existing text, match the density and register of surrounding paragraphs.
- Direct and conversational register, but no contractions in formal writing. Personal essays and conversational pieces keep their contractions; the no-contraction rule applies to formal and technical writing only.
"Turn this outline/lose idea for an article/4 paragraphs of text into a blog post similar to these previous blog posts, but make sure that this one has a table of contents and a bunch of references"
It reads that way and Pangram says it's AI. And my experience says that if you see an AI-related headline on HN, there's a 50%+ chance that it is AI-generated and meant mostly for clicks.
>Have you tried putting known human writing into pangram? I have. I've gotten 100% AI with multiple samples of my own human writing. It has also given me 50% on things I know were 100% AI written (from my prompts).
>Pangram is basically a made-up number. / I've tried it on large docs I've written well before the AI times, and that are nowhere available on the Internet (so it can't be a corpus issue) - and it is happily classifying me as 60%-80% AI.
Unfortunate there's an incentive to pay to sign up to protect oneself against false accusations.
An earlier claim in this thread stated 100% from the same tool, but another commenter claims 76%, so apparently the tool is even susceptible to that failure mode.
Skepticism against AI text detectors on HN is as old as time, and frequently comes from people with some vested interest in filling up the internet with slop when you look at their business ideas / projects / blogs. I've done systematic testing on human and LLM-generated text and I'm confident that the accuracy is higher than 95% (and that <5% is almost exclusively false negatives).
You shouldn't crucify people based on this alone, but if it reads like AI, quacks like AI, and is detected as AI, it's probably AI.
>Skepticism against AI text detectors on HN is as old as time
Since early 2023 or so, when the detectors were widely reported (off platform) as unreliable?
>and frequently comes from people with some vested interest in filling up the internet with slop
I'm sure sometimes.
At least once, has come from someone who recently philosophized about whether to call out the specifics of AI writing in the first place given the potential for aiding labs in their training missions https://news.ycombinator.com/item?id=48326913
>I've done systematic testing
You in the field?
>but if it reads like AI
Actually didn't to me, and I'd like to think my detector's no worse than the average for a commenter here... perhaps as we'd all :)
The article didn't do a good job explaining the 120% attention angle, I kept reading waiting for that and it never really came. I definitely had the impression it was heavily using AI in the writing which I gave up on being against, but it just didn't explain the thesis well.
I guess the idea is AI gives you back time so you could now do the 20% but you still really can't because you have to still think about it even if the code is generated? Not even sure after reading all that text what the idea is
.
Regarding the the current HN post's title, I like the TFA original title ("The New 20% Time, Minus the Time") better, along its sub-title; IMHO, the current HN post's title ("Google's 20% 'project' has become AI's 120% 'attention'") is more like reflection of the HN submitter's interpretation of the content of the blog post (that I agree with partially and to a small/negligible extent ONLY).
---
IMO, keeping "Google 20% Time" (mentioned in the TFA) and "R&D is two jobs, ..." (his second-last blog post, at time of writing this comment) in mind while reading this blog post, helps in reading and understanding the content of this blog post; whether we agree (or not) with the author's point-of-view is another matter.
---
Btw, I also noticed AI usage on this blog post; however, I over-looked that part of the post after looking at author's past work (and I'm happy I did that and continued reading the post).
I never worked at EA, but they had "Friday Afternoon Project" at least someone who worked for EA here in Florida told me so. The unfortunate thing about Friday Afternoon Project being its shorthand acronym or "F.A.P." not sure if that was intentional or just a "happy accident" but a coworker found one such Friday Afternoon Project and had sent me a youtube of it, it was pretty funny looking, thing of a really bare bones game concept basically. I guess it was a way for EA to let employees have some downtime.
I was always jealous of the 20% off concept, because there's so many jobs and places where I'd use that time to solve things nobody wants to "fund" within my org, sometimes there's some really dumb bug somewhere, or easy to solve for internal tooling need (I'm sure Google has had this resolved many a time internally) that could be met if I could even have two hours on a Friday to work on anything.
The other problem is that R&D used to be taxed differently for decades, then people abused it and ruined it for everyone else, making it so only FAANG level companies can afford R&D.
Artists and actors could point the LLMs at programmers by developing their own apps, doing their own engineering. Do they? Who's doing it already? Tilly Norwood's creator seems like a decent start.
Distracted by the LLM-generated style of writing. Not sure whether the author truly writes like that (unlikely) or there was heavy AI assistance in drafting this.
What prompt would you use to generate this? I don't see it.
I have no idea about the current article, and given that the author is the person with the first commit in the Kubernetes repo (https://joe.dev/about/), he obviously has a lot of credibility.
Just generally though: what we're seeing a ton of these days is people writing something and then passing it to an LLM with a request to improve it somehow, e.g. by fixing grammar, tightening the style, etc. In such cases, the answer to your question is that the "prompt" is (1) a first draft, and (2) an instruction to edit it.
It's clear, though, that the LLMs leave far more imprints on the text than most people realize, and that although they may have asked the LLM to restrict its edits to "just" X or Y, the actual changes to the text will often go beyond that.
How this will evolve over time is anyone's guess, of course.
I have a steering .md file that instructs Opus how not to sound like an LLM when writing prose (I write prose in my IDE with Opus). The steering is specific to me, but I've found that giving Opus rules like eschewing punchy journalistic sentences ("Not because X. But because Y. And that matters."), varying sentence lengths and avoiding staccato sounding clauses go a long way in smoothing out LLM smells in writing (at least according to me).
Aside: different LLMs sound different too! ChatGPT is the worst offender for LLM-sounding writing and needs the most smoothing, but Claude (web) actually sounds like a humanities major from the get-go.
I do this so many times. Type in a large amount of text and the only thing final in my mind is para breaks and the idea per paragraph. And then give it to AI saying "sending to director", "sending to friends on WhatsApp group", "sending to colleagues" and it does an awesome job of bringing the "AI polish" and then you edit or negotiate line by line or para by para on what you want to keep.
Yes, this is certainly common, and opinions and tastes differ about the outcomes—as they should, because we are all still in the early stage of sorting through the best way to use these tools. I think it's also already clear that "best way" means something different in different contexts.
Where some people are getting into trouble, at least in the HN context, is underestimating the impact that this has on their text. There's a big perception gap between the author's view ("fixed up the grammar a bit") and the reader's view ("this sounds entirely like an AI wrote it") in many cases. So many, in fact, that I feel I can say something about it. I'm no authority on any of this and don't want to sound like one, but this is such a common pattern at the moment that I feel confident reporting it. How it will change over time, I have no idea.
(I also don't want to sound anti-LLM - we rely on these tools heavily, they're amazing, they've already improved HN, and they show every sign of high potential to improve it further. The bottleneck isn't the LLMs, it's how quickly we can figure out how to use (and test) them. We just don't use them to process any text that we put on HN itself.)
Attached is the writing.md I use to steer Opus 4.8. Prompt:
I ran it on TFA and Pangram flagged it as LLM generated but Claude Fable couldn't definitively tell.-- writing.md ---
# Writing Rules (MANDATORY)
## Banned Words and Phrases
Never use: "incredibly", "extremely", "absolutely", "fundamentally", "dramatically", "crucial", "vital", "powerful", "robust", "elegant", "seamless", "cutting-edge", "game-changing", "groundbreaking", "It's worth noting", "Importantly,", "Interestingly,", "Let's dive in", "At its core,", "At the end of the day,".
No exclamation marks in technical writing. No contractions in formal writing.
## Banned Sentence Structures
1. *Semicolons joining independent clauses.* Do not write "X does A; Y does B." Use a comma + conjunction that names the relationship: ", while" (contrast), ", and" (addition), ", so" (consequence). Semicolons hide the logical link and sound artificially balanced. 2. *Label-colon-explanation.* Do not write "The key insight: ..." or "The limitation: ...". State the point directly or use "is that" phrasing. 3. *Colon after a bolded term.* Do not write "a *rollout engine*: a lightweight...". Use a comma. 4. *Sentence fragments as assertions.* Every claim needs a subject and verb. "No gap at any ρ." → "There is no gap at any ρ." 5. *Em-dashes joining independent clauses.* Do not write "X does A — Y does B." Use a comma + conjunction. Parenthetical em-dashes ("the policy — trained offline — cannot adapt") are fine. 6. *Tricolon lists of near-synonyms.* "It does not X, Y, or Z" is padding unless each item is genuinely distinct.
## Banned Rhythms
1. *Staccato sequences.* Two or more consecutive short declarative sentences of similar length. Join them with a conjunction or subordinate one. A single short sentence standing alone for emphasis is fine and often good. Do not eliminate it. 2. *Formulaic layout.* Do not produce: intro paragraph → three bullets → summary paragraph. 3. *Gratuitous parallelism.* Do not force list items into identical grammatical form if it makes them sound robotic. 4. *Saying it twice.* If you stated a fact, do not rephrase it from another angle in the same paragraph. One clear statement is enough. 5. *The negation-correction reversal.* This is the move where you deny one candidate and assert the real one. Surface forms to match: "not X, but Y"; "it isn't X, it's Y"; "X was never the point, Y was"; "for me X, for them Y"; the comma-tag "Y, not X" ("sanctioned, not stolen"); the "not so much X as Y" form; and the gapped version where a stranded verb delivers the pivot ("Hours aren't the bottleneck. Attention is."). One reversal at a genuine turning point is good writing. The structure is not the problem. The density is.
6. *Repeated hedge adverbs.* A softener like "almost", "somewhat", "rather", "a bit", or "fairly" used more than once in close range becomes a tic. Keep at most one, and only where it earns its place.## Positive Rules
- Active voice. Use "we" and "our". - Concrete nouns and verbs. "The model overfits after 50 epochs" not "exhibits suboptimal generalization characteristics." - Plain English. Use technical terms only when they carry meaning plain English cannot. - State consequences, not meta-commentary. "The policy has no lookahead" not "training compresses multi-period consequences into a single-step mapping." - One sentence that advances to the next thought beats two sentences restating the current thought. - State assumptions when uncertain. Do not hedge-stack ("it might be the case that perhaps..."). - Not every paragraph needs a topic sentence or a concluding sentence. - Do not resolve the ending with a tidy bow. A piece may close on an open question, an admission, or an unresolved tension. Summary endings that restate the thesis read as machine-generated. - Do not over-smooth. Removing every short sentence, every parallel, and every fragment flattens prose into uniform medium-length sentences, which is itself an LLM smell. Fixing a tell should not cost the voice. - When editing existing text, match the density and register of surrounding paragraphs. - Direct and conversational register, but no contractions in formal writing. Personal essays and conversational pieces keep their contractions; the no-contraction rule applies to formal and technical writing only.
"Turn this outline/lose idea for an article/4 paragraphs of text into a blog post similar to these previous blog posts, but make sure that this one has a table of contents and a bunch of references"
It reads that way and Pangram says it's AI. And my experience says that if you see an AI-related headline on HN, there's a 50%+ chance that it is AI-generated and meant mostly for clicks.
Recent Pangram feedback on HN:
>Have you tried putting known human writing into pangram? I have. I've gotten 100% AI with multiple samples of my own human writing. It has also given me 50% on things I know were 100% AI written (from my prompts).
https://news.ycombinator.com/item?id=48326698
>Pangram is basically a made-up number. / I've tried it on large docs I've written well before the AI times, and that are nowhere available on the Internet (so it can't be a corpus issue) - and it is happily classifying me as 60%-80% AI.
https://news.ycombinator.com/item?id=48378226
Two of my own thoughts:
Unfortunate there's an incentive to pay to sign up to protect oneself against false accusations.
An earlier claim in this thread stated 100% from the same tool, but another commenter claims 76%, so apparently the tool is even susceptible to that failure mode.
Skepticism against AI text detectors on HN is as old as time, and frequently comes from people with some vested interest in filling up the internet with slop when you look at their business ideas / projects / blogs. I've done systematic testing on human and LLM-generated text and I'm confident that the accuracy is higher than 95% (and that <5% is almost exclusively false negatives).
You shouldn't crucify people based on this alone, but if it reads like AI, quacks like AI, and is detected as AI, it's probably AI.
>Skepticism against AI text detectors on HN is as old as time
Since early 2023 or so, when the detectors were widely reported (off platform) as unreliable?
>and frequently comes from people with some vested interest in filling up the internet with slop
I'm sure sometimes.
At least once, has come from someone who recently philosophized about whether to call out the specifics of AI writing in the first place given the potential for aiding labs in their training missions https://news.ycombinator.com/item?id=48326913
>I've done systematic testing
You in the field?
>but if it reads like AI
Actually didn't to me, and I'd like to think my detector's no worse than the average for a commenter here... perhaps as we'd all :)
Is this more AI writing, or are my sensors just on the fritz ?
Pangram reports 76% of this text is AI generated.
Did I miss news about an authoritative writing analyzer?
[dead]
The article didn't do a good job explaining the 120% attention angle, I kept reading waiting for that and it never really came. I definitely had the impression it was heavily using AI in the writing which I gave up on being against, but it just didn't explain the thesis well.
I guess the idea is AI gives you back time so you could now do the 20% but you still really can't because you have to still think about it even if the code is generated? Not even sure after reading all that text what the idea is .
Regarding the the current HN post's title, I like the TFA original title ("The New 20% Time, Minus the Time") better, along its sub-title; IMHO, the current HN post's title ("Google's 20% 'project' has become AI's 120% 'attention'") is more like reflection of the HN submitter's interpretation of the content of the blog post (that I agree with partially and to a small/negligible extent ONLY).
---
IMO, keeping "Google 20% Time" (mentioned in the TFA) and "R&D is two jobs, ..." (his second-last blog post, at time of writing this comment) in mind while reading this blog post, helps in reading and understanding the content of this blog post; whether we agree (or not) with the author's point-of-view is another matter.
---
Btw, I also noticed AI usage on this blog post; however, I over-looked that part of the post after looking at author's past work (and I'm happy I did that and continued reading the post).
I never worked at EA, but they had "Friday Afternoon Project" at least someone who worked for EA here in Florida told me so. The unfortunate thing about Friday Afternoon Project being its shorthand acronym or "F.A.P." not sure if that was intentional or just a "happy accident" but a coworker found one such Friday Afternoon Project and had sent me a youtube of it, it was pretty funny looking, thing of a really bare bones game concept basically. I guess it was a way for EA to let employees have some downtime.
I was always jealous of the 20% off concept, because there's so many jobs and places where I'd use that time to solve things nobody wants to "fund" within my org, sometimes there's some really dumb bug somewhere, or easy to solve for internal tooling need (I'm sure Google has had this resolved many a time internally) that could be met if I could even have two hours on a Friday to work on anything.
20% time works when the 80% time spent is a money printer that goes brrrr
only a few companies like google had that imo. most companies cannot afford that.
The other problem is that R&D used to be taxed differently for decades, then people abused it and ruined it for everyone else, making it so only FAANG level companies can afford R&D.
I worry the answer to "new surplus. who will it go to?" is a foregone conclusion. the surplus is the most surveillable technology imaginable.
Artists and actors could point the LLMs at programmers by developing their own apps, doing their own engineering. Do they? Who's doing it already? Tilly Norwood's creator seems like a decent start.