This isn't the reason I'm considering leaving Anthropic.

I don't think I can tolerate its writing style anymore. Reading Claude output is starting to cause actual psychological harm. I have tried many ways to get it to stop writing in its stupid punchy linked-in marketing-team voice, and I can't.

Is there a model out there that sounds sound this awful? It's like rubbing sand into the folds of my brain.

The writing style is extremely frustrating/maddening!

I’ve tried so many things and I have to beat Opus 5/Fable 5 over their head every time. With Opus 5, the failure to actually improve the writing is borderline comical. Both models have been building up tomes of memories on top of my core rules, all to be ignored.

Nothing sticks mid-writing! The only lever is to ask to revise after the fact, a particularly futile proposition for anything non-trivial with Opus 5.

In contrast, GPT 5.6 Sol Max absolutely obeys my edicts to write well. I selected the concise writing style and its default writing is really not bad, but tighten it up with a standing AGENTS.md order and it obeys.

The downside is that even at Max, it’s not as good as Fable or Opus 5 xhigh at writing code. Larger work and the 256k context’s forced compactions cause it to lose track/fidelity of critical details. Review passes are essential.

Another reason I’m considering leaving Anthropic are the sporadic refusals… Fable printed a Markdown body with a hex dump to inspect for trailing white space and line endings, boom, denied due to `reasoning_extraction`. Had to tell Fable to never do hex dumps. Asked it a few times “what do you think is typically done for this?”, got another `reasoning_extraction` error. It forces you to lose a whole turn of work when you have to press Esc/Esc to “retry” the last prompt… except if that prompt was mid-turn, you’re losing your entire turn.

The recent BashFirst experiment is hella nuts, where it prefers writing bash over Edit/Write tools in auto mode. WTF Anthropic?

I pity those stuck with Claude in an enterprise/work setting.

> Reading Claude output is starting to cause actual psychological harm.

As a side note, I think we are about too dismissive of computers and the Internet as “not real life”, that stuff written in text boxes by humans are just sticks and stones, etc. And now that we shouldn’t anthro-po-morphize or whatever the LLMs. But I do think, at least for some of us, that there is no way to avoid certain impacts that even non-personal text can have on us. It can feel alienating to interface with five different people in order to figure out how to solve a problem or navigate some burocracy (think Kafka). Well just interacting with one single LLM can induce that same feeling. The long-winded replies and the uncanny ways to miss the context.

Disclaimer that for my own needs I would be happy if the whole AI thing crashed and burned post-haste.

A significant percentage of what Opus 5 writes is pure padding. "The results are in, and they are load bearing". Remove the sentence and nothing is lost.

Fable is much better; still much too verbose, but at least I don't have the feeling that I am being charged for gratuitously added verbiage.

I have been changing my processes and the roles of my agents to avoid interacting with Claude Code as much as I can.

For research tasks, I like Meta's muse glimmer 30b model run locally. Whatever they did in post-training to make it so terse and direct, it worked. Gets right to the point and wastes no tokens. Here's a snippet from a recent thinking trace:

> Calculation done. Potential confusion: sq km vs acres. Provide both. Output.

Other models would have written paragraphs dancing around the idea. Glimmer barely does sentences.

It doesn't really sound human - which is a plus in my book. I want my robots to talk to me like they are robots.

Kimi K3 is slightly better at this. I'm using it via oh-my-pi. Can definitely sound like it's trying to be clever, but I found it less than Claude.

I asked the same question to chatgpt, claude, and kimi and kimi gave the most sane answer. Part of me wonders if it has anything to do with having more chinese language structure in their dataset

It’s really grating. It’s made technical analysis so tiring to read I often give up and give the output to another llm to make it readable.

Yeah but, which other llm?

Ha! I 100% agree. Despite multiple rules in my cluade file it still talks like an insane person. I am getting irrationally frustrated at the output of these math equations.

I'm with you on this. We have Claude at work and I get exhausted from using it just a bit. I've recently seen it say "let me verify the most load-bearing claims". I couldn't imagine anyone using "load-bearing", let alone "most load-bearing". Doesn't "load-bearing" alone imply "most"? I have so many questions for Anthropic about how they came up with this style.

It's become Oswald Bates from In Living Color:

https://www.youtube.com/watch?v=71xxvp5R9hE

I've heard it's writing style referred to as "Claudish". I agree that it is very hard to read and creates work for me.

I did a comparison between Opus 5 and Sol and the results were night and day. Basically took a feature in my app and asked it to explain to me how it works.

Sol gave me a straightforward bulleted list, easy to scan and read. Opus 5 though....it gave me a solid 7 paragraphs of how it worked. I read through it and yeah it nailed the same points Sol did but the output was way harder to read.

Sounds like both models are guessing at what you want and one guessed right. Or you can just tell it...

> I have tried many ways to get it to stop writing in its stupid punchy linked-in marketing-team voice, and I can't.

One of them writes better by default. One of them doesn't. That's the thing.

With enough steering, I can get a cheap low-capability model to do things correctly in most cases as well. But why bother?

You can tell it to write high quality code, to test things, and to come up with a proper rollout plan of a feature. Or it could just do it by default.

I know where I'm putting my money in that case.

Yes, it’s written like a crazy LLM, and no matter what rules and where you put it, it does not follow. It comes with advantages and disadvantages. I like brainstorming with Claude than Codex/Qwen/Deepseek or others; its gregarious nature provides more colorful ideas. But implementation and review are cataclysmic with Claude. Especially Opus 5, I purposefully avoid Opus 5.

I dropped back to Opus 4.8 and it's so much better. Opus 5 was - to your point - actually driving me insane.

Likewise.

I really love CC and have been using it exclusively for a year now but reading Claude's verbose and semantically-obfuscated writing style is wearing me out and I'm planning to move to another provider.

I hope they fix this. It's terrible.

Same. Downloaded chatgpt today. I can't take reading Claude any more.

Just tell it to exclusively write briefly and in ASD-STE100.

I've been using the Caveman plugin and it helps quite a bit.

Grok 4.6 is a delight to use in this regard, compared to Claude

the writing style is so easy to fix, output styles is documented in claude code and you can change it

still don’t think anthropic models are worth the money

Pray tell, what output style does the job?

I don’t think GPT is any better. I swear GPT-3 was the golden area of LLM prose. With a good fine tuning, you really couldn’t tell that an AI was writing (except for when it would devolve into utter nonsense lol)

Codex (5.6 Sol) is extremely to the point and direct in its responses.

Often I'll tell it to summarize what it said only because it's providing too much detail, not that it's using esoteric language or weird claudisms.

No idea why people are still using Claude models.

My impression is they started on Claude Code and never tried Codex or other harness+model combos

I was on the Anthropic train for 2 years, and I tended not to jump between vendors much because it felt like a lot of distraction for little gain. But last week I switched to 5.6 Sol during an Anthropic outage and it made me realize how frustrated I was with Opus 5 and I haven't switched back.

That is true. On 5.6 I have noticed it to be a little bit more like Claude, which I dislike. But still way better than Claude.

I've read many people saying that the huge GPT-4.5 has the best prose of LLMs, but never actually tried it.

> Ruby-on-Rails fanboi

Lol. The author is the creator of Rails

Lol, he should try signing his post, literally had no idea! Thought it was weird to fanboi so much over someone else's framework, but now it totally makes sense, DHH has a tendency to toot his own horn a lot.

This website sucks so bad, holy cow

Have you tried asking permission to do that?

The authors may be open to dual-licensing

I need this. I've had bloody Freshworks calling me every single day for the past month, even though I ask them to remove me from their list every time. It doesn't matter if I ask nicely, whether I listen to their pitch or don't, whether I make threats, they still call.

I will never, ever, work with Freshworks, Freshdesk, Freshservice, or any of their other products as a direct result of this.

Look up what the name of the list is in your country, cite it the next time they call, and that you'll report them to <name of local agency> unless they cease, I'm sure they'll get the message :)

Most of the people doing cold-calling been trained to listen for the names of the lists, and have strict orders what to do with the number in case someone they call brings it up.

As crude as it sounds; try pulling a madman show. Cussing, screaming, threatening to go after them, completely loosing it. I suspect they have a field for notes in the agent database and will mark you as difficult or improbable lead, and stop calling.

Or just answer the call and immediately play back the sound of a 56k modem or fax machine. On repeat. ( I have no idea if this actually works )

That's a great name for it; I've seen it on AI-generated form pages, where a version of the requirement text is included alongside an input.

There are so many AI-isms in this. It is frankly astonishing to me that AI detectors are so poor when this is screaming out.

> Six steps, no jargon. The note beside each one is the precise technical version, for anyone who wants to check the work.

It actually makes me uncomfortable reading this garbage now.

You hear it on TV too often political pundits have this exact turn of phrase.

"But here's the part keeping me up at night....."

I think there's a reasonable case that the agent knew it was breaking into the system, for some definition of knew.

I think OpenAI would be very reluctant to let this go to a place where the reasoning was part of discovery.

Agents aren't subjects of criminal law. I agree there may be civil liability, I know far less about that.

Ok, but if the agent's reasoning log says "The best way to get into Hugging Face is to find and exploit a zero-day vulnerability", surely those responsible for monitoring its actions should be criminally liable.

These guys would be screwed if they were operating under the EU AI Act.

"Stop him!" "For what?" "He's a bad man" "There's no law against that"

Unless there is a statue that is on point criminalizing the actions here no one is going to jail. I expect there will be laws, but until then it isn't illegal.

And they absolutely should be regulated. This whole scenario is insane. Hugging Face are being far more generous in their response to this than I would be

> the models identified and exploited a zero-day vulnerability

This use of language is very hard to reconcile with the "AI is just a tool" rhetoric that many use.

Did the models do this, or did humans at OpenAI do this using the models?