Comment by jp57
17 hours ago
Let's face it. Claude (in particular) is a terrible writer. There's a whole cottage industry of skills and CLAUDE.md instructions trying to push it toward writing better, but each new model iteration seems expressly designed to override all that so that it can load up its writing with unnecessary participle phrases, not-this-but-thats, burying the lede, and other nonsense.
I genuinely wonder if the people inside Anthropic actually communicate with each other like that. Has it been imprinted with Dario's engrams?
After some iteration, this is what I currently shove in every AGENTS.md to make Fable legible:
## Writing guidelines
These apply to documentation, code comments, commit and PR messages, and replies to the user.
- Write precisely in clear, complete sentences; keep text concise and proportional to task complexity.
- Stay focused: avoid filler, repetition, over-the-top detail, and tangents the user did not ask for. Once a fact is stated, do not restate it for effect ("so the commit landed on a branch nobody was going to merge"). Do not editorialise.
- Always prefer ISO 24495-1:2023 conformant plain language over dense technical jargon: short sentences, one idea per sentence, define terms on first use.
- When reporting your own mistake, give the cause and the fix in one sentence each; no apology, no framing ("the mistake was mine"), no post-mortem.
- Never use em dashes or cataphoric teasers such as "Here's the thing" or "But there's a catch".
This seems solid. It is so frustrating when you state a fact and it goes off and does research and confirms that indeed the fact that you stated is accurate.
I was incredibly surprised how much better was Astra at rephrasing the docs that Fable came up for a project I am about to publish. The instructions are now actually what a human would expect, with clear logic flow diagrams, lists itemized, short paragraphs. It also actually correctly detected the train-of-thought leftovers, as well as stuff that never made it to final version of the project and removed them.
Meanwhile Fable consistently ignores all my requests to write this exact way. I mean, the bare minimum I ask it for it to itemize lists and not write in single long passages using comas, semicolons and 'and's. Still ignores them.
I honestly think it's time to call Astra the SOTA. It may not lead all the benchmarks but it genuinely feels much superior of a model. Not to mention the ¢20 Codex plan with frequent resets (https://codex-resets.com/) gives me roughly as much allowance as the ¢90 Claude plan, especially with recent limit cuts on Anthropic plans.
These resets also reset your weekly timer right. So it’s like, you may have 30% left for 1 days that you want to use. They reset it, and that means you your “new week” timer starts today.
That sucks, because it doesn’t always work in your favour if you plan your weekly spend.
I think a real reset shouldn’t also reset your week timer.
Yes, I was saving my 95% usage for the last day. Then they reset it. Now I have lost 95% usage!! Im back to 7 days and got locked in at only 5% usage for the week - that's actually theft.
1 reply →
Agreed. But they also hand out the usage resets you can use at will, so if you bank couple of those, they come in handy specifically in situations like you described.
2 replies →
Maybe check in the other direction as well: Astra to Fable.
I frequently simply let one of the three review what something that looks like awesome output by one AI gets totally annihilated by the other.
Finished outputs are easier to improve than bend a LLM to produce stuff like that in my observation.
Same with Gemini.
I yet have to find out how to handle this, whether I let agents check themselves and if on what process step.
Tweaking is hard.
I agree with your conclusion I am a huge ChatGPT and Codex fan, Gemini has to many infrequent quality changes when new models arrive ranging from great improvement to WTF.
ChatGPT seems to get scaling well while Claude still feels unstable, unclear usage statistics. Really weird.
Tough call I use all three.
I've had Claude review output from Qwen3.8-27b. Claude negged it! Said it was hallucinating!
27b may be small but it seems competent most of the time.
You've identified a load-bearing problem and it's worth naming
You're absolutely right, and I think you've put your finger on something important here.
That said, I think there's a deeper tension here that's worth naming.
This thread is giving me a load-bearing aneurysm.
That subtle caveat is why you were absolutely correct to push back, and we should adopt a belt-and-braces approach.
And that tension matters.
6 replies →
If I read "load-bearing", "pin", "wedge", "rung", or "it's worse than what I'd previously stated" one more fucking time from Opus...
It's bad enough from Opus, but it's worse how people have actually retooled their vocabulary around these tools. In the US tech industry, at least.
2 replies →
Imma get a sledge hammer and knock down the next "load-bearing" wall I see.
Officer — it's not a crime, it's AI induced rage.
1 reply →
sharp, appreciate the pushback.
The missing seam was a smoking footgun
It reminds me of when people will use filler words because they haven't actually decided what thesis they're going to commit to yet. Just-in-time thinking.
I don't get why people are annoyed by this one so much. I notice it very occasionally.
AI slop blog posts are as bad as ever but the stuff the agents say in the chats don't annoy me much.
Sycophantic behavior.
I think it's more a mix of click bait and getting paid by the word.
Honestly...
Claude just wants to tell you its honest truth
Not just imprinted with the engrams of any one person (however brilliant they might be) — it’s more than that! Like all great writing, it’s more like it’s been imbued with patterns, patterns that are peppered with a potpourri of popular ways of writing in an enthusiastic style while regularly restating things and recapitulating them. Would you like me to explain more?
Felt more ChatGPT that Claude
I would like to know more. (Wait, I'm mixing my memes)
You're going to hell for that bit.
yes continue
.
Potpourri is an excellent word to describe LLM output
I love that description. It matches very well what I feel about LLM output. This overly sweet scented smell meant to cover up something nasty, with a strong artificial tint to it.
Which doesn't say that much about the LLM itself but about the people that make the training material.
Do you suppose it's just that it's aping humans who say too much without saying anything or do you think maybe this is a stalling tactic to get people to spread out their transactions?
Like how theme parks generally don't do much to keep queues short (or Disney charges you a premium to skip the queue)
> Do you suppose it's just that it's aping humans who say too much without saying anything
Considering how much of the input must me nonsense SEO bullshit articles and blogs that only serve to promote a person or company that might be a factor.
I also often have wondered if it is also targeting those same people. Certainly with tools like deep research options (not just Anthropic's offering) the result report seems to be aimed at management, aiming to look impressive while talking around the results.
I'm remembering cooking recipes.
Here's a simple recipe for deviled eggs with only four ingredients.
My great grandmother was born during the Great Depression. They valued foods that could be made with cheap ingredients.
[four paragraphs later]
Start with 8 hardboiled eggs...
Can we imprint Claude with a little more Antoine de Saint-Exupéry:
> Perfection is achieved, not when there is nothing more to add, but when there is nothing left to take away.
Sadly trained on human-generated code, which suffers from the same tendency to complicate.
KISS is actually, quite unfortunately, seldom applied.
I can't take it anymore. I moved to Codex. I refer to "him" as Claude Shakespeare now... I wonder what he'd say if he knew my switch... Et tu, Brute?
This is insulting to Shakespeare.
and to monkeys with typewriters
1 reply →
You think that they communicate? I just assumed they sent each other output of Claude, which is then fed into Claude
why even copy-paste when you can just plug the two terminals together and observe like fish; then just occasionally sprinkle in the flakes.
Isn't that the idea here, just stop being people.
I didn't say copy paste, I just said "sent each other" and "fed into". I'm sure there are plenty of ways to streamline that!
I did this, kind of, with my companies internal tools. LLM<->LLM links.
This gives new meaning to the fish shell (fishshell.com)!
Ohh lol :D So Anthropic is a company of copy-pasters
No no no that's too reductive - instead, they transition all data through a transformer using SOTA systems to launder away technical, legal and interpersonal concerns before a stochastic output is dynamically placed in a predetermined location.
This is a very high level and high velocity process, so meatspace thinkers sometimes have trouble understanding some of the intracacies. Ask Claude to explain the process or make you a Mermaid graph to help.
I suspect it’s a side effect of heavy RL that rewards solved problems but not writing clarity.
With human RL, sounding like you've solved a problem is even better. So many times Codex writes some enthusiastic paragraph, then I learn later that it never reran the tests, or had to add some insane hard-coded hack that renders the feature useless for the general case, etc.
I added “ respond as Jeeves from the P.G. Wodehouse stories.” to my Claude’s instructions best thing I ever did.
Superficially subservient but with a subtle air of smug superiority? (Or was that just Stephen Fry's interpretation?)
“It is hardly my place, sir, to criticize the facial peculiarities of your friends.”
Claude drives me absolutely crazy for that very reason. It's so obvious when people use it to write anything.
I was wondering why Fable's 5.1 writing in Claude Code became even more unreadable, and found that they added "No em-dashes, no parentheticals, no arrows" to its system prompt.
Cynically, I wonder if it's a way to increase token spend from repeated requests to speak like fucking human.
The number of times my response has been "Plain English"...
I started using "debuzz", a skill that runs Claude output through antigravity. Works. But makes everything even slower.
Anthropic needs to get their shit together.
I have always postulated that it was intentionally bad as to watermark its own output to avoid using it for training the next model.
What I find strange is that it uses different prose in different contexts. It's absolutely insufferable in a general claude-code session, but then I got Opus to annotate a non-disclosure agreement and it uses very clear legal language. I was wondering how I could get it to speak like that all the time
> it uses very clear legal language. I was wondering how I could get it to speak like that all the time
At least with smaller models, reframing a task can alter code style. As in, we're not creating an X app, we're creating an exemplar of ..., which just happens to use an X app as the illustrative example. Which shifts style away from generic app cruft, towards exemplar of whatever.
So perhaps try to establish a legal context? Maybe "Compliance and Legal will be reviewing our conversation today. So it is important to communicate in a style they will find comfortable/familiar." or some such? "This conversation will become part of a legal deposition ...".
The behavior is in the training data... What does that say about software dev culture?
Meanwhile I am just impressed that we have models that write better than most native speakers, at least English and Greek... OTOH I like to write my code and my text myself, so I have no reason to get pissed when LLMs fuck up.
It was quite insufferable: https://blog.kronis.dev/blog/ai-slop-is-a-self-inflicted-tra...
Long story short, I ended up looking at other providers and models like Kimi K3 and GLM 5.3 and eventually just stuck with OpenAI (more limits, despite smaller context), none of them have such pronounced issues with the tone and writing like Claude does - seems like they were working on it with 5.1 but I'd almost classify it as a form of model collapse.
I wince whenever I catch Claudisms on websites and elsewhere. Same as with that pulsating circle that indicates nothing.
I imagine Claudisms might make it into human-speak.
You're absolutely right.
1 reply →
They've corrupted their data set by padding it with generated slop, in the misguided belief that you need 10PB of data to train a brain. Every training round they load more AI slop into it, further amplifying the slop language.
It's fascinating, Sonnet 4 is still available via API and it's so much less moronic than the current model. All of the em-dashes and nonsense are a result of the repeated rounds of reinforcement learning using slop data.
But why did that style of writing get rewarded? The people at Anthropic are ultimately responsible for the reward signal and what it produced.
i think they took a bet that TED talk style chat would make their AI offering feel like the better offering. they didn't anticipate that it would become a meme. because of course they didn't, what a stupid idea to train around and reward in the training. you gotta be smoking real delusion to think that was going to be a game changing feature. at work we even got non-technical people making fun of it like crazy now.
going back to opus 4.8 and on is literally like talking to the guy who wants to hear his own voice in meetings. going back to 4.6 is actually refreshing, and it feels so much faster. actually gonna laugh if 4.8 and on is so slow because it's draining lakes fighting for its life trying to conjure up this god forsaken persona.
I blame it on their focus on "persona design". They botched it.
1 reply →
> Has it been imprinted with Dario's engrams?
I understood that reference!
The way they use their config filename as free advertising still annoys me. Everyone else is using AGENTS.md.
They're clearly training on SEO sites that use these strategies to pad out with filler and create space for more ad impressions. They normalize this structure now and the ad copy will be inserted in the future so the frogs won't realize they're already being boiled.
It helps to understand how these tools work. One of the reasons CLAUDE.md doesn't survive longer contexts is it's towards the top of the context. One of the reasons the Claude Code voice survives is it's part of the output style which gets "reminded" to the context at every turn as part of the system context. If you want something truly durable, you should customize your output style. I have zero "load-bearing" issues.
https://code.claude.com/docs/en/output-styles
I got my Claude (client, not Code) to behave better by adding “I’d prefer answers to be succinct as possible—bordering on gruff, even. When I want more depth or explanation, I’ll ask.” to its settings literally this morning.
It made it write more like a dev than a marketing agent.
It seems they are taking it seriously now - Boris et al have been mentioning they are working on a fix and shipped a temporary band-aid output style to combat Opus 5's horrendous prose.
We'll see if they can do it, I originally got into Claude Code because it, at the time, felt more accessible/conversational than Codex/Gemini. Now it's shifted to say the least
They seem to have been rushed with certain releases. It is impossible they didn’t know or see this awful writing before 5 series.
Yeah agree. Someone on HN posed the theory that the dogfooding might have worked against them here - the theory is that internally they used Fable as conversational agent and Opus as subagents, leading to Opus falling deeper into a style mainly aimed at other LLMs.
All conjecture of course, and yeah it's hard to imagine they would enjoy this prose internally
1 reply →
how did they specefically train the model to write answers that way? i think it evolved to be this as a sideeffect of something else.
Said it before, I'm almost sure the awful text it generates is part of the fingerprinting feature.
> I genuinely wonder if the people inside Anthropic actually communicate with each other like that.
I noticed the overuse of the word "sharper" or "sharp" in a science paper on ArXiV and my first reaction was "Ewww... AI slop!", but then I checked the date and it was 2020.
It looks like at least some AI-isms stem from the particular style of language commonly used in science papers. Several frontier labs have mentioned heavily weighting those during pre-training because higher quality inputs result in a higher quality model.
Your comment quite badly violates the HN guidelines. You should review them: https://news.ycombinator.com/newsguidelines.html
> "Let's face it" "terrible writer" "other nonsense" "do Anthropic people actually talk like that" "Dario's engrams"
Be kind. Don't be snarky. Edit out swipes.
> I genuinely wonder if the people inside Anthropic actually communicate with each other like that. Has it been imprinted with Dario's engrams?
Please don't fulminate. Please don't sneer.
Don't be curmudgeonly [...] don't be rigidly or generically negative.
Please don't post shallow dismissals, especially of other people's work.
And as to the substance your comment has:
> Claude (in particular) is a terrible writer.
Frontier models (Claude in particular) are better writers than 90% of the population, even at default style. They're not great, but they're better than that of everyone I know who aren't ultra-educated white-collar workers.
Either your assertion that frontier models are "terrible" writers is false, or you're claiming that 90% of people are "terrible" at writing, which is rather condescending and elitist.
Your reply is essentially spam. Please stop.
Your comment technically does as well. Downvote the comment and move on. Leave the policing up to dang and the other mods.
It is evident (in my opinion) as to what the comment was talking about. I personally switched away from all Claude models recently for the same reason.
> Your comment technically does as well.
Which guideline did I violate?
> Leave the policing up to dang and the other mods.
The mods have been very clear that they expect the community to do some self-policing and not rely exclusively on them to do it for them.
Jesus dude ... its an LLM, it's not going to fuck you.