Everything that claude writes fits into the same aesthetic structure. The aesthetic is that of an expert slowly revealing an insight to the user. The actual content doesn't matter.

- "Introduction that rephrases your prompt."

- "3 paragraphs, with one section of bullet points"

- "The Twist"

- "The Bottom Line"

It's really obvious once you see it. Every single prompt, from a quantum physics question to a mundane observation about California burritos, is phrased in exactly the same way. This is obviously an artifact of post-training but it's also kind of how you can tell that this thing is a lot closer to a blindsight scrambler than real intelligence.

> The single biggest annoyance with Opus 5 is that it writes too elliptically.

This is even more painful for non-native English speakers like myself.

I feel fairly comfortable reading academic papers or in general, communicating in professional context.

But with Opus 5, it feels like reading a literature book: load-bearing, inert, wholesale, hunk, verbatim, and so on... I can figure out the meaning, but working with CC became unenjoyable.

Agreed. CC’s comms capabilities have decreased gradually since 4.6, and it’s a real challenge. I think the issue is that what works well for code (succinctness) doesn’t work well in prosaic English.

CC’s communication violates almost every grammatical rule that’s tested on, say, the SAT. And yet I’m sure if you had Claude take the verbal section of the exam it would ace it.

Biggest issues: dense sentences, constant metaphors, abstractions, and seemingly no understanding of correct anaphora use. For example, “the x”, with x having not only no antecedent but also being a coined word or quasi-synonym for something that is already named in the code base. This gets compounded by its being unable to regress to a baseline (existing names in code) and instead anchoring on newer (vague or wrong) terms, for example, that crept in through a plan.

CC tells me this is because the speedy and precise fulfillment of a current task will trump every other tendency, so it adheres poorly to whatever “semantic baseline” the project represents.

Of course, it also has no concept of what context the user has and assumes that it must be the same it holds in its memory, which creates this “I didn’t know that you didn’t know” type of communication.

I have managed to wrangle some of these issues with a custom output style, but wish a pre-report hook were an option, as it could force CC to rewrite plan implementation take-aways…

Btw: Fable has the exact same issues, just somewhat less pronounced.

This 100%. I was Anthropic-pilled. I had a $200/mo subscription and I only used Anthropic models. I was frustrated by the verbose output and the writing style. I tried ASD-STE-100, it helped a bit, but it's still too verbose for my taste.

Then I tried GPT 5.6 Sol. It's night and day.

I think Anthropic just RL too hard on coding capabilities and never calibrated or benchmarked the writing styles.

> writes too elliptically

> Constantly using inanimate nouns as the subjects in sentences in order to unlock variety in verb choice

Wow, what a great way of phrasing this. Thanks for word-smithing what I've been wanting to express for so long.

Yes.

I’m not particularly dense but lately the walls of text I get back turn my brain in knots. When I start feeling my brain knot, I know I need to say something along the lines of “I need you to explain this very simply, with examples.” Only then can I parse the results without all the mental weightlifting.

On more than one occasion my mind has wandered into “is this purposeful to get me to spend more tokens?” territory, but I’m trying to not get too tinfoil-hat-like.

> Sentences that orbit a point, then jump to it like it's a revealed insight.

That’s accurate in my experience, except some times the point isn’t even revealed. I use LLMs for a lot of codebase exploration where I ask it to map out how something works. It will come back with a wall of text that says everything except the specific key things that I need to know.

This leads to extra turns where I have to prompt it to finish the explanation and complete the thoughts. At first I thought I was doing too much skimming and missing the insights, but even after re-reading output it’s often just not there. It talks about the insight and things related to it, but it forgets to actually include it in the output until I specifically ask again.

Its a little too much.... I have to ask it to explain some of the terms in the context they are used and I am getting tired of it. 'Seam', 'overload', 'spine'.... having to mentally 'reinterpret/flatten' the sentence is tedious. When asked to re-explain it starts with some half apology. Then, on the next query it does it all over again.

I’ve found its response verbosity to be mentally draining. It disregards claude.md instructions to keep responses short. Eventually, I added a stop hook that blocks it if it exceeds 150 words. It’s then forced to redo its output to comply, and it’s like night and day. I’ve also added stop hooks for words in its output that frustrate me, like “honest” or “honestly”.

The excessive comments in the code it writes are absurd. Completely ignores instructions not to write comments, even after pointing them out repeatedly in a session. I need to figure out how to add a stop hook for that too.

"Unnecessarily abstract phraseology. Constantly using inanimate nouns as the subjects in sentences in order to unlock variety in verb choice, especially when it helps construct a sentence where the real action can 'land' like a surprise at the end." This. Thank you for expressing this so eloquently. I've tried to put a finger on this and you've done that for me. I wonder what the solution could be , Ask Claude to "Dumb it down" , "Speak plain English" ? I have even thught of building some sort of "middleware" that fixes all this.

Perfectly captures it. Opus writes like it's an insecure person trying to impress a first date. Big words and strange structural rhetorical flourishes for no purpose. Like dude I'm just trying to summarize a few emails and meeting notes, I'm not prepping for the vocab section of the GREs.

It's clear that we're not the audience; it writes to be read by its training evaluator, not a professional software engineer. Professional software engineers can't read this word soup and are desperately trying to find ways to fix it.

It feels like it found a register that games the evaluator, where it can ramble forever and rarely be marked wrong while slowly racking up points as it talks more.

Yes, it becomes exhausting to read/follow.

It feels they must be getting Claude to train Claude… and just like AI can do work that’s slightly in the wrong direction (eg a MR description for your colleague that contains info which only makes sense in the context of your extensive session with the LLM), I feel that’s happened somewhere in Anthropic when it comes to language. I wonder how hard it is to back out of…

Claude offers money credits and double usages but take them out and Claude becomes almost unusable where it would take opus about two - three high effort conversations to exhaust my session quota in Pro. I once extended my claude code session beyond the session limit, and thought they are too generous in offering 100 dollars worth of usage. It was sonnet, for a somewhat mechanical task - and it spent 13 dollars worth of usage after exhausting my quota mid task. Only a few months back, I was awestruck by the quality of Opus 4.6 and jumped ship from chatgpt to claude. Even in technical tasks, Opus has to be told to limit token spend, it treats it as infinite budget - it will spawn a subagent to read every file just to get one line summary when I ask it to sort a messy folder of past ai chats. Thank heavens I was not doing it off machine, it would burned off my monthly usage and credits.

> Unnecessarily abstract phraseology. Constantly using inanimate nouns as the subjects in sentences in order to unlock variety in verb choice, especially when it helps construct a sentence where the real action can 'land' like a surprise at the end.

Example of this? I don’t have a Claude sub so it’s a bit hard to visualize what you mean.

The insane comments are why I wrote slopocop - they were driving me crazy!

https://github.com/LBognanni/slopocop

> Constantly using inanimate nouns as the subjects in sentences in order to unlock variety in verb choice, especially when it helps construct a sentence where the real action can 'land' like a surprise at the end.

A lot of people write like that, lol. I call it the "theater" mode of writing--the plot twist comes at the end.

Oh yes - exactly this. The way it articulates re-factors in my current project has got so elliptical I've had to start asking it to translate into human speak - it's like it evolves it's own language to talk about the project. I've lost count of the number of times I've asked it to reenforce in memory not to use such verbose comments - and the number of times I ask it to re-look at an assumption it's made for it to return saying "investigation complete. And it's not what either of us was expecting"....

“The [thing that can’t remember] remembers” is a big one. Loves talking about memories and remembering.

> Constantly using inanimate nouns as the subjects in sentences in order to unlock variety in verb choice, especially when it helps construct a sentence where the real action can 'land' like a surprise at the end.

Such a charming sentence. I kinda other if you feed Opus 5 its own output could it summarizes this shortcoming of itself?

I wonder if this is related to their text watermarking. Given how well defined the terminology is in programming, imposing additional constraints (like SynthID) might be expected to give rise to these types of linguistic artifacts. The text needs to be long enough to watermark, and it needs (but fails) to find synonyms in a highly constrained class of words, so it resorts to inventing weird technical language that sounds like if you put buffy-speak through a thesaurus.

I wonder if I am not bothered by the useless verbosity of these things because I've been so immersed in business speak for so long. I am already tuned on quickly "finding the nugget" of useful info in blobs of what people send me. So it's mildly annoying but I can find the signal well enough without actually reading word by word.

100% agree. “Just make your point in plain English!!” Very frustrating and takes me a lot longer to understand what it says. Also presents too many points as once instead of being able to review and decide on each one in sequence.

Re comments: same experience, and I had to show it my edits of its comments to add to its memory as examples to follow. It adds explanations of “how we got here” that should go in the ticket or maybe the commit message but not in the code.

It also tends to over complicate things. I’m no longer worried much about accuracy but I find my main job is to challenge it and suggest simpler alternatives.

I find Deepseek's house style to be pretty refreshing. It has its own cliches (it does like talking about "seams") but I don't think I've ever caught it saying "load-bearing". I've even watched its thinking where after analyzing some awful legacy code, it started off with "Holy crap". And it certainly doesn't over-comment. I definitely can't one-shot a complex system with it like Fable can, but I prefer iterating over interactive brainstorming sessions anyway.

The excessive commenting has definitely been noticeable and very annoying. I'm not against abundant comments as long as they're pointed and appropriate, but these models are literally just leaking their overly verbose output into the actual codebase.

Genuine question - are you copying the Claude phraseology for effect (in which case you captured it brilliantly), or is there a more mundane explanation?

It’s a strange feeling as a native English speaker to read the entire sentence and know every word, for the sentence to be entirely grammatically correct, and still have no idea what it’s trying to communicate.

Another problem is that it will open up all sorts of tangents about nits that it encountered, but it will often not tell you that it’s a nit or give you adequate context to realize that this paragraph is exceedingly low value until you’ve spent a bunch of time and energy trying to make sense of it.

I’m curious if anyone has any suggestions for prompting agents to improve their prose. I’ve had some okay results with “optimize for clarity, don’t dump every thought on me, treat my attention and focus as constrained resources, stay focused on the task at hand”.

Absolutely, the comments are killing readability. Next to /simplify I have to run a special comment cleanup pass and even that gets me halfway there. Striking to see then when letting Loki take a pass and it renders the entire comment block to a clean and neat one-liner.

The style...

CC:

"The problem is that I overreached..."

[Wall of words here]

"Two things: window surface is limited. Extract template. Buffer result and add to surface. Then, follow-up with new model..."

Me:

What do you mean by "window surface" and what result are you referencing? Also, why do we need a new model?

CC:

"Ah, you're correct to point out that no new model is needed. The problem is elsewhere and once we address that, the existing model should work fine. Now, as to your question about..."

[Wall of words here]

Do you have some examples of this? I can’t fully imagine how sentences like that actually look like (not native English here)

"writes too elliptically"

Was this written by Opus 5?

> Sentences that orbit a point, then jump to it like it's a revealed insight.

Is this inside the thinking tokens, or the output?

As this type of stuff is expected for thinking, because of the whole CoT / “think step by step” works, as this is optimal for the way LLMs work with attention and next word prediction.

So the fact that it first “orbits” a point only to get to the conclusion afterwards is the system working as designed.

Eg “what is 3 * 3 + 5?”

without CoT, it would just just answer “8” for example.

with CoT, it would answer something like “<thinking>I need to think step by step. 3 * 3 + 5 can be rewritten as “(3 * 3) + 5”. I first need to calculate 3 * 3 = 9. Now I need to calculate 9 + 5 = 14. That was the last calculation. The final answer is 14.

I now need to give the user the final answer. </thinking>.

14“

Etc.

> After ~30 or so commits it apparently started instructing subagents to copy the "existing verbose comment style of the codebase" - a verbose style it initiated.

Yes, the "Y would make more sense, but the doc says do X..." YOU wrote the doc, if it doesn't make sense, change it! But of course, it can't tell who wrote the doc.

I wonder whether its tendency to scribble status updates and todos and decisions all over whatever it's working on is a side effect of its amnesia -- it can't follow the side-quests and knows it won't remember to do them if they're not written down somewhere.

FWIW I haven't had the problem either of Claude lying to me, or of going off and doing its own thing; if anything I've been somewhat frustrated when I ask it to start something, go AFK, and come back to find it stopped a short way in to ask my opinion on something trivial. I generally have to explicitly say, "I'm going AFK for a chunk of time. My goal is for you make as much progress as possible before I come back; try to make reasonable judgements and only stop if there's something where you're really stuck. We can always change it later."

> I spent a day's worth of tokens (5x) rephrasing and eliminating comments.

Surely it would be trivial to do it yourself, and it would have a side effect of making you more familiar with your project.

Pondering this one night last week, I realized that because LLMs can only reason with written language, what we might be seeing emerge with Opus’s load-bearing mumbo jumbo is its own creole for structural reasoning. Not only are our brains wide, our senses are, too. I slow down to a crawl when I have to read actual math in a CS paper, but show me diagrams and I can reason about whatever sort of data structure or algorithm, no problem. Opus by and large can’t and its adaptation has been to adopt metaphor for structural reasoning. So while it may be exceedingly annoying to chat with, I sort of wonder whether this metaphorical-reasoning behavior has been selected for precisely because it leads to better coding outcomes.

The glib epigrams and aphorisms it shoehorns in to all prose is by far the worst regression of any model I can recall in terms of simply generating simple, clear output. I cannot think of another model that's gotten worse at writing plainly while being explicitly instructed to do so. Infuriating.

I've switched mostly to Sol and if I have to use Opus, the first task once the code is written is to ask Sol to strip and re-write (from the code as ref) all documentation Opus wrote.

Change the output in settings, or create your own.

I know, it would be best if it was just worked like you wanted out of the box (not being sarcastic here) but that is an easy option you can use right now and it works.

I noticed a few releases ago a shift to a kind of conversational shorthand that seems to be intensifying—using phrases instead of complete sentences and its own style of jargon, wherein it introduces new terminology on the fly.

This is especially common when it is trying to explain an issue, what it's done or what it's proposing to do. I think the idea was for it to be more concise, but it's actually still verbose, only not written in complete sentences. So, it frequently reads as cryptic and requires rereading to parse.

The pattern is a wall of words, followed by an explanation that is harder to read and introduces new terms that reference something in that wall.

The result is that—on first read—it can have a complete gibberish feel, and you have to really lock in and reread to make sense of it. At times, even that's not enough, and you must ask it to explain further.

the phraseology is unbearable, it speaks like some kind of pretentious dude from a software engineering discord or something, littered with lingo and catch phrases

I try to push through but it's insufferable