One thing I've noticed and HATE, is that when you increase thinking-effort, that seemingly increases response-length. Meaning that X.High is longer than High, which is longer than Medium, etc.

Which is kind of the inverse of how people work; a really smart person can condense difficult ideas into simple[r] terms. Whereas people who struggle speak a lot but say very little.

High/X.High do seem to deliver better quality results, but it sometimes feels like needle-in-haystack extracting that from the word vomit.

With LLMs, you're still mostly read things "off the tip of the tongue". A better comparison is observing a smart person talking to themselves while working on a tough problem.

EDIT: also there's a reason the dial is called "effort", not "smarts".

Yes, but that should apply to the CoT "thinking", not the final output.

> They clear their mind and let the solution come.

They work on a problem until their brain is full of problem-related concepts. Then something comes. After validation it might be a solution.

It's so bad I've made myself a Pi extension that rewrites responses in side by side view using models on Cerebras (insanely fast tps)

It's so bad I made my own chat client for Claude, so I can attach steering prompts in conversation. They are applied just at the end, before the last LLM response, then removed and response kept.

I just go over the comments with Gemini 3.1 Pro at the end which has a much more normal "voice" and it doesn't lose nuance as a cheap model would. I don't care so much about what Claude writes during the debugging as I just do all the cleanup at the end instead of at every commit.

The higher the effort the more things Claude checks, and it's eager to tell you about all of them

See, this insight it had early on looked like a red hering for a while, but then turned out to be load-bearing. And that's not just a difference in semantics, it changed the whole conclusion (spoiler: it didn't). And Claude is very eager to tell you about this exciting journey

“I have made this letter longer only because I didn’t have the time to make it shorter.” - Blaise Pascal

> when you increase thinking-effort, that seemingly increases response-length

I use the /sss writing style - synthetic, short and simple - and it helps a lot.

OpenAI has separate dials for verbosity and reasoning_effort (but could still do a better job).

I hate this too, I had to switch to Codex, because the skill to force Claude Code not to think too much about very, very basic things no longer worked

I just can't stand how often Claude says something like "And the honest part? It's..."

Like, were the other parts not honest? I don't understand how Anthropic let it get like this, it's been such a clear regression

Sometimes Opus 5 (high/xhigh) feels like I'm dealing with the programmer equivalent of Zeno of Elea.

Every time, without fail, it would get me 90% of the way there and then leave a small note, exception, or deferral. When instructed to address that, Opus would somehow take nearly the same amount of time as the first 90%. And then it would finish with yet another deferral. Repeat ad infinitum.

You can sometimes get around it using the `goal` directive provided you are not subject to the constraints of mortality.

They got that from Anime seasons. Every prompt has yet another cliffhanger to keep you hooked. But the Season II story arc where Claude-chan fights the NsPasteboard boss battle on the journey to the UIViewMainController, I thought that was pretty intense. I guess I just gotta keep watching my terminal to see what happens to the main character input - rooting for him to survive the next season, but you know they always kill off the good input characters early.

Yes and the last bit is always mysterious and inscrutable. I have to think way too hard to figure out what the actual problem is. I’ve noticed it does a lot of explaining the mechanics of the problem it found, but almost never explains why it’s important until I ask.

And the worst part is that this little problem will keep sneaking into the context of future sessions, unless you spend the time to fix it. Even if it isn’t important, I’ll sometimes have Claude fix it so it will shut the F up about it going forward.

I am often asking it to write in sss-style - "synthetic, short and simple style"

i think they took a huge bet that speaking like a ted talk was going to be a vast popular differentiator in their offering, i don't think they anticipated that people were going to make fun of it, that it could become a meme..that it could get in the way of getting stuff done and result in cancellations.

it's downright exhausting to read claude, the language style was a regression imo.

If you ask it why it uses the term honest so much it'll tell you it was actually trained not to. lol

It can't truthfully answer "why" questions, only infer them in a way that aligns with its training for conversational engagement.

Of course, that's how LLMs work. It's still amusing, at least to me.

Geminis is "it really is". The Notebook podcasters use it _constantly_.

I wonder if I can make a tool for it to write messages back to me, say that it can only speak to the user through tool use, and then put a hook on that tool to prevent any of the Claude-isms

Me too. And it does it so often, that I've added a stop hook that detects "honest*" in its response and forces it to regenerate without the banned word.

I think it says "honestly" when breaking bad news.

"Humans have a token limit too" - that's so good and it explains so much of the fatigue that myself and colleagues/peers have about Claude in particular.

I think it's not just token limits - I think it's because it's so _dense_.

You get a week of research and debugging and testing compressed into a few pages. Even if it's explained well, it's just so much information. And since it's AI, I'm constantly second guessing "is that really true?" and it's exhausting.

I've developed a habit of adding into my prompts "please keep your response concise and succinct" or "I'm trying to cram, please only provide the minimum level of technical detail necessary to understand this topic"

I find it helps immensely but it'd be nice if I didn't have to do that.

i tried using claude codes output style option to do something like this and it worked for like three prompts and then it was back to normal lol

I don't understand that complaint, although it seems to be a common one. The whole problem with the way models talk nowadays is that they are succinct to a fault, going to the extent of coining new buzzwords and misusing existing ones. What I want to see is a shift towards plain language.

I’ve had good results adding “also I’m a baby” to the end of all my requests for explanations

"I am a thickie thickie dum dum" works even better. Plus it is a good reminder.

Memento moroni!

ELI5 always works

Sounds like a good way to get even more refusals, LOL.

why so many people add 'please' when asking machine to do something? Was there actually research that when you SCREAM or curse it follows your instructions better?

P.S. Although my wife insists that I should stay polite in case AI overlords remember how I treat them ...

I'm polite to LLMs. It's not for the models it's for myself. If I start being rude to models then I might accidentally start being rude to other people as well.

Probably because polite people are already in the habit of saying please when typing out requests in chat. We're not consciously thinking about it, regardless of whether a human or machine is on the other side.

Not to go all ying/yang about it, but just to give a parallel: https://en.wikipedia.org/wiki/Loudness_war - you kinda need silence to draw a contrast with what's meant to be loud.

Separately, my boss confided in us that he's super abusive with his agent, wondering if we are too (no, lol). While I try not to read too much into this (which he doesn't make easy), I also can't help but not really notice a whole lot of amazing agentic delivery differences from his side. On the contrary, while the passion may improve his agent's performance, I'm not sure if it doesn't decrease his, upending the entire theatre.

I think about removing please/thanks, but then I accidentally add them back in during some edit/rewrite of the prompt... It's just how I'm used to asking for things

Or "be brief and simple"

you could apply it on lifecycle hook level, probably the most appropriate place for it

add to your system prompt?

It only sticks to the instruction for maybe 3-4 turns. This is why when Anthropic released "concise output style" feature in claude code, it basically spams the model's context with "be concise" system reminders every other turn.

ya, but now i have this system that ive build that works with any ai agents, putting boundaries, gates every single time and it generate memories from the runs so it can inject them as needed.

just try kaplira, it will help u with that.

> Prose density is not the same thing as succinctness

Can't agree with you more. I review 2-3 PRs a day from my team of eight data engineers. Most of my team members use Claude to write SQL, dbt and Python code. Some of them use Claude a lot, some less so. I can easily tell when I review the code that is mostly Claude generated vs. the one that is not. In dbt models where we have a lot of biz logic in intermediate layers, that's where I really have a difficult time following Claude-generated comments. So much jargon copied over from other adjacent dbt models (yet inconsistently), and the prose is super choppy (for the lack of better word).

After reading a looooong sentence/comment line, I still can't figure out what it really means. Had to always re-read the line 2-3 times (sometimes, more) to sort of understand. Reading code, however, is so much easier and usually, I just skip to reading the code and then come back to the comments. :D

Amen. I would trade some stupidity (say ten points on any benchmark) in exchange for a version of Opus or a similar model that actually gave me direct, concise answers.

You should try setting claude code to opus 4.6. With the style instructions I set in my user CLAUDE.md it does exactly that. It's like night and day: Opus 5 gave me a page and a half of word-vomit, yet the exact same task and prompt with 4.6 and I got maybe 100-150 words total, entirely readable.

x2 on opus 4.6. still works great, and it's fast. opus 4.6 is where i hope local llm's get to someday, that's kind of my personal benchmark for where "local is more than good enough i dont need these idiot large-scale service providers"

Yes, and they will work ... for like two turns, after which Claude will go back to its usual wall of text.

And yes you could add context (memories, rules, CLAUDE.md entries, etc.): they won't help (for long). Same for hooks that remind Claude to be concise: it gets "attenuated" and starts ignoring any such instructions quickly. There's also writing guidelines ... but they're basically just more context with slightly higher weights (ie. Claude will still ignore them).

I've even gone so far as to make a hook that identifies long responses and requests shorter versions (which is challenging in itself, as you need to run another lower-powered model to evaluate how long is "too long", as what's "long" when the expected answer is one line is different from what's expected for a ten line answer). However, that just shows you the long version, then some hook text, then (10-15 seconds later) it shows the short version. So I created a proxy that hid the long version/hook text for me ... but I had to abandon it because all that used up so much usage I was running out.

I'm fuzzy on the details, but Caveman somehow "hacks" Claude in a way that gets past all that ... but it takes things too far in that direction, with "cave man" speech that sucks.

I switched to using Codex for the last two weeks, and while the prose has been better, there have been a lot more technical oversights. I'm now having fable review codex commits and it finds deep issues. I'v also done the reverse where opus/fable do the work and then I have codex revise all of the prose prior to reading anything myself. This has also been effective; I'm not sure which is the better approach.

Just the other way I was thinking that if I asked "What does Lamborghini do?" the only correct way to answer is a single sentence "Which Lamborghini are you referring to?".

But LLMs will fail at this question: they will tell you about Lamborghini's latest car and mix some history in it. Just try.

Which is the wrong answer anyway, because there's at least two major companies called Lamborghini, one making cars, one making agricultural equipment and at least one famous person (Elettra) with that family name.

This very simple test/question makes me realize how much do I hate LLMs in a sense: while I agree that the answer it gives is the most plausible for 90% of the users, it's ultimately both wrong and long. And that 90% compounds.

But there's no "correct" answer in my eyes than "who are you referring to?". Possibly without listing all the possible Lamborghinis.

This is ... unnecessarily pedantic. Anybody in my social universe who asked me that question would undoubtedly expect "they make cars".

If you're picking nits, why not focus on the word "do" and (wrongly) expect an answer like "Lamborghini (either of the two main companies of that name) does not 'do' anything - the companies employ humans who 'do' things. Lamborghini is a legal entity established to allow humans to 'do' things, such as make cars, or agricultural equipment."

Shared context is a thing. Reducing every conversation to first principles is not always required. Get a grip.

This is not a nit. This is a real problem in the technology.

It assumes the average and plausible answer token by token.

And this tendency shows in every single field it's applied to.

At the end of the day I want *correct* answers, to the point.

Instead LLMs, no matter if it's version 3500, are bound to producing average results: slop.

My biggest frustration with Anthropic with Opus being too verbose is that they tried to put this on users. It’s pretty clear that Anthropic employees don’t use the day-to-day models that everybody else use. They have access to the next tier model so they don’t see the problems that everybody else is dealing with.

Yep, they have no clue what their users are complaining about since all they use all day is mythos max preview.

I wasted a lot of tokens last month asking "Please explain the meaning of this sentence in plain language"

Kimi K3 is the best explainer. I switch to that for code review and insights.

You can change CC's output style (https://code.claude.com/docs/en/output-styles). You can also put style notes in your global claude.md. I've instructed claude to treat me like I have adhd, get to the point, and be succinct, ... More or less eliminates the problematic prose.

I took time to figure this out after Fable spat out "...then stays purely as cascade-debugging provenance rather than load-bearing arbitration."

My experience with output styles for long-running sessions is that Claude starts to forget the terse output style by the middle of the context window. Obviously I don't know if 5.1 suffers the same fate but I ran into this issue with both Opus and Fable 5

That sentence is fine; it’s tolerably annoying. As a long-time HN reader, HN is full of this kind of performative erudition and I’m already used to it. Fable probably learned from the worst parts of HN.

I just used it to do a review of a ~100k SLOC codebase that Fable 5 / Opus 5 largely built, cost like $2 and caught some good stuff, but more importantly, it communicated very directly and was pretty light on bizarre metaphors. No "let me read the source before opining" type verbiage launched at me. Honestly night and day for me vs before.

If you ask any model to write as tables to enumerate points, and BDD for logical flows, it’s like 50x less strain on you

Same, currently on a mix of Kimi Vivace (K3), GLM Max (5.3 and 5.3 Flash) and OpenAI Max (Sol and Terra mostly).

I will say that Kimi feels nice but slow, GLM feels faster but has limited tokens (even off-peak) and OpenAI is nice and fast but has limited context (258k shows up in Codex, really).

Neither of them are perfect, but I prefer their type of prose across the board to what Opus 5 and Fable 5 kept outputting. I'll probably check out Anthropic again in a year, but for now I need a break from its brand of slop. Oh also all of the other ones allow usage in OpenCode with their subscription plans.

> In some cases, though, its prose is denser than Claude Fable 5's: sentences run longer and there are fewer paragraph breaks.”

This sentence reads like Claude wrote it. Perhaps it did, or perhaps Claude has learned to write like the folks who work at Anthropic?

(Had I edited this, I would have said that a colon is not the right separator here. The second clause does not _explain_ the first, per se, bur instead expands upon it. Consider instead: "In some cases, however, its prose is denser than Claude Fable 5's, with longer sentences and fewer paragraph breaks.")

Going off of vibes, I guess this would call for a semicolon or an em-dash?

Also, could be just Claude rubbing off on them than it being Claude authored. I'd imagine they read it quite a bit.

Today, Opus talked about "rotation slabs" in relation to logging. (and not log rotation). I didn't even bother asking what that was supposed to mean and switched over to Sonnet.

Just remember Charles Dickens was paid by the word too

It still talks the same claudish, but now it's indeed denser. I'm not quite sure what step up they're talking about.

Same here. I still have access until my account churns but Anthropic has huge issues comparative to everyone else with token / usage burn down. K3 Swarm also delivers better results than Fable at a fraction of utilization. The Pro plan is definitely not worth it anymore and if I do want to burn some money I can always just leverage the API. But Anthropic went from simply amazing last year to a dumpster fire in less than 6 months for my use cases, anyway.

> sentences run longer and there are fewer paragraph breaks.

Gotta fit in the watermarking.

Yes! I have my .md's have

"If you respond with more than 3 paragraphs, give me a TLDR"

"Do not assume I know all technical jargon, please explain things plainly"

> In some cases, though, its prose is denser than Claude Fable 5's: sentences run longer and there are fewer paragraph breaks

that feels like they just blocked words like load-bearing but can't actually fix the real problem. The insane word slop density and run on sentences was the real reason it became annoying to work with claude, colored with way too many analogies and pointless linguistic comparisons.

I don't think they care about humans... ... ...

You know you can just... change this right? What a weird thing to cancel over. I have in my global CLAUDE.md something to the effect of I have ADHD and give me succinct responses with headings and lists etc, works great.