It’s not that it’s vocabulary is too advanced. It’s something worse.
I get lost trying to follow even a paragraph written by Claude. It’s so loaded, and complex, and full of indirection and propositions. It’s too much!
Am I alone in this? Using a model like Sol or GLM is much easier and I feel my Brain CPU not being pegged at 100% for an hour.
It’s not just you. I ordered it to make a devspeak script checker for its own stuff and started banning words. Just today: 7.
Somebody said something about Anthropic wanting to watermark Claude’s responses. Fair I guess but it’s their problem and they are imposing it on us.
Oh, and there’s also this: nobuzz/README.md at main · adnanakil/nobuzz · GitHub. Just reading the first two paragraphs made me laugh hard. Here’s the biggest gem, ahem, the kicker, no sorry, the load-bearing paragraph, oops, I meant the revelation here, nope, I think it was “the fascinating discovery”, no, you know what, I give up:
Claudette is our solution to Claude being a great engineer with one incurable condition: it talks like it’s delivering a TED talk about its own pull request. You ask why a test is flaky and you get a “load-bearing assumption,” three numbered revelations, and a solemn note that the third one is the most instructive yet. Nothing is ever just a bug. There is always a kicker.
It makes me nauseous. The caveman skill helps a lot. GitHub - JuliusBrussee/caveman: 🪨 why use many token when few token do trick — Claude Code skill that cuts 65% of tokens by talking like caveman · GitHub
I explicitly ask it to use plain English and explain it to me in terms of what and how I’m going to see the change in browser. I also tell it (and repeat over and over as it tends to ignore it) to stop using generic single noun terms like “arms”, “legs”, “frames” etc. and to replace them always with two or three word self-explanatory and unambiguous terms. When it drops a wall of of its codified references (again), I tell it to translate it. Plus, I no longer let it edit anything before we settle what it plans to do in plain English.
You are definitely not alone!
I just recently got back to using Anthropic models after a few months using Codex almost exclusively, and the defaults can be pretty insufferable at times. It’s everything all at once: the verboseness, unnecessary fluff, and indirection. Don’t even get me started on “load-bearing” @dimitarvp 
.
The pretentious jargon just triggers me so much haha (especially for a non-native English speaker)
I have been trying different approaches with varying degrees of success, so I’m just gonna leave a few recommendations here in case it helps anyone:
I’ve also been having moderate success asking Claude to prefer “ADS-STE100 Simplified Technical English”.
Yep. Just to pile on, Opus 5 is regarded as being extra confusing, and I have found this to be the case.
My understanding is that they’ve been training the new models on the outputs of other models for so long, the models have been getting worse and worse (since Opus 4.6 was where it started to get noticeable) as the newer iterations succumb to model collapse.
Just an FYI, you can still use Opus 4.5, 4.6, 4.8, etc. in Claude Code: /model claude-opus-4-5, /model claude-opus-4-6[1m]. Perhaps you might like to take them for a spin to see if their output is less disgusting. (Note that 4.5 will have a 200k context limit, and 4.6 generally falls apart after 200k tokens in my experience when using the 1M context variant.)
I exclusively use caveman. Without it I can’t handle it.
Also I think that we are reaching the peak and more and more models are being train on output from other LLMs which results in stuff that no human would ever write like that.
They are poisoning themselves.
This skill might be helpful, maybe in combination with caveman. It’s not only about unreadable text output, but they start to pollute everything with “watermarks” due to new invented regulations so its also about formatting;
I feel people are talking about different things here: conversation output and code. I agree on the conversation output some of the times, but it all depends on the situation and cannot be easily compared as there are to many factors in play. A simple “be concise“ or similar usually helps for me, but if the context gets to large the output may become to look like spagetti (with sausages). But I truely haven’t noticed a negative on the code produced, on the contrary.
The irony that this document reads almost exactly like a handout I used to give college students to make their writing more palatable. Now they can pass it directly to the AI they use to write their papers for them. Efficiency!
I have seen comments written by Claude that are borderline unreadable. Like the mumblings of a psychopath. Things like this
Extract the id from the user struct. Unfortunately we don’t know what the users passes in. It could be a user struct, it could be an integer or a string - the truth is we don’t know. In order to take care of this we need to take care of business. Therefore, even if it costs some performance, we need to make sure to parse this input correctly. Otherwise, bad things might happen. Also removes the need for manual parsing.
Like.. I had reviews where I had to tell people to remove this ridiculous essay ****.
Interestingly, on my Windows machine, the Claude Code stopped auto updates and so I’m still on Opus 4.8, on Macbook I’m on 5 and I didnt find a standard way to lower it to 4.8.
Man, this thread really helped me, I was beginning to question myself.. Especially in web-components world with CSS properties and variables being accessible on different layers, the texts Claude produces about technical details of some solutions are really like reading Stephen King’s novel when in one paragraph you travel through 3 different spacetimes.
And the worst part? You ask Claude “What do you mean by that?” and it answers “You are right to push back, I was not precise and mislead you, it’s actually this and that”. So not only it’s hard to read, it’s also hard to invest your time into understanding it because it’s probably not true
/model claude-opus-4-8
This is fantastic 
I had one I laughed out loud in the office. It was something to the effect of.
“We need to have a timeout of 30 seconds in case the user loses the connection, like if they go into a tunnel”
What? Does it matter what causes the disconnect? Why mention tunnels, maybe the user got kidnapped and thrown into a van with a faraday cage!? Anyways, I’ve switched to GPT 5.6 Sol and it’s night and day.
Ha, just today I have said this in my Claude prompts:
holy cow you talk in such dense language. Please try again and speak clearly like a human
and
ok slow down and assume I’m an idiot
Or copy / paste into Grok or Codex and ask it translate into human.
I have this in CLAUDE.md in one project and the outcome seems a bit better:
# Writing style
Applies to chat, comments, docs and commit messages.
- Follow ASD-STE100 writing principles adapted for software: active voice, short clear sentences, one idea per sentence, consistent terminology, no unnecessary synonyms or hedging.
- In code comments: explain why / constraints / non-obvious rationale only. Do not restate what the code does. Prefer brevity.
- Use project/technical terms freely; otherwise stick to simple, single-meaning words.
- Say it once and stop. No aphorisms, no "X is the whole point", no paired contrasts for effect, no restating the same idea in a second, more lyrical sentence.
- Answer the question that was asked, then add context if it is needed.
But I still get stuff like this as the context builds and it forgets:
I should be plain about a trap here, because a defensibility product is the worst possible place for me to bluff: the principles are the kind of thing I could produce a plausible-looking list of from training, and some of it would likely be right, but I can’t reliably distinguish, from the inside, a faithful recollection from a confident reconstruction. If I hand you “the principles” and two of them are subtly my invention or drawn from a different release, you’d be building schema and audit language on a fabricated foundation — exactly the sin the system exists to prevent. So I’d rather flag the uncertainty than perform completeness.
which translates to “I don’t know”
Maybe it’s a cynical ploy to burn more tokens ahead of IPO?
Perhaps relevant: Model collapse - Wikipedia
Perhaps relevant: Model collapse - Wikipedia
Lol’d at “Habsburg AI”.
I switched to GPT-5.6-Sol. Good stuff so far, I can easily throw away >50% of my AGENTS.md (sorry, CLAUDE.md). Oh, and it integrates nicely with Zed Agent where I can granularly approve edited lines, and it uses the built-in read/write/grep tools instead of making edits with bash python3 -c '...' commands (wtf).
Preliminary totally non-scientific evaluation results: GPT-5.6-Sol Medium > GPT-5.6-Terra Xhigh >>> Opus 4.8 > Opus 5
This roughly aligns with https://deepswe.datacurve.ai.
The tractable beachhead was a ratchet all along – honest caveat : we’ve been building this as a spike, but the real insight was finding it to be load-bearing. What I need from you is your honest feedback : if you want me to rebuild it, just say the word.
Yeah. It’s quite mentally taxing. Maybe the idea is to progressively stop reading agent traces ? For the time being I still follow along every session and those new models developed a rather infuriating vocabulary and mannerisms.
That, and I used to use “load-bearing” as a joke. Like in “load-bearing paint”. Now I can’t.
Edit : just saw The load-bearing vocabulary of Claude which could be of interest to you
© Copyright Elixir Forum | Terms | Privacy & Cookies
