Comment by karimf
3 days ago
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.
Yeah I don't know that any of the benchmarks index on "understandability". I'm amazed at how Claude can produce a page of text describing what it did and it can take me a full five minutes to decipher it, often just to find it's something I could have expressed in a simple sentence.
I just spent a day writing very thorough system prompts for communicating in different contexts.
Everything is super succinct. Opus 5 lands, it almost completely disregards the intent.
I suppose watermarking requires a certain text mass.
The watermarking is going to get rolled back or Anthropic is going to get rolled. People hate it and it makes the writing worse.
1 reply →
Oh man. Hadn't even considered the watermarking angle.
2 replies →
They want you to use Sonnet to explain what Opus is trying to say. They're not optimizing for token efficiency.
Adjust the output in settings. Or customize it to what you want.
Have you tried asking it for a lay explanation of what it did? That’s usually all it takes for me. Sends garbage -> request -> sends something readable
Brilliant way to get people to waste tokens.
Maybe just don’t generate garbage in the first place?
No, I’m not interested in fighting my model all day long. Plus is fucking annoying to talk to and collaborate with, so I’m not using it when Sol 5.6 is about 1000 times better in that regard. I have colleagues who spent a lot of time trying to improve their harness with user rules and whatnot and Opus really does not want to follow them.
Yeah but Sol shows it is possible to just send the readable explanation in the first instance. And I don't want to spend tokens and time on asking for a better version of each response.
When I ask it to make a CL description, it's worthless unless I tell it to dumb it down as much as possible, assume the reader has zero knowledge of the codebase. And then it makes a perfectly cromulent description that just needs a touch of trimming-down. If I don't do this, the description is just a wall of gibberish and paraphrasing of every little thing it encountered.
Yeah my trick is "Restate concisely"
Just those two words. I use it A LOT recently.
It's a surprising change from my perspective, because in the past it felt like they understood that Claude should be pleasant to interact with.
It's bad enough that I've seen dedicated skills to do comment hygiene scrubbing and consolidation.
I've tried telling it to "fix" comments with varying degrees of specificity and in my experience it just... fundamentally doesn't get it. Presumably using a different model for it would help.
My theory is that Claude's learned approach to comments is to treat them as a sort of persistent in-band thinking trace, or a "memory" tied to an in-code location, which is a little at odds with the way humans use comments (human comments are intended to be read and understood by other humans, whereas Claude comments are their own dialect).
I bet this is a result of iteratively training Claude on output from other successful Claude sessions. Presumably it's good for making benchmark scores go up.
9 replies →
I built my own skill to somewhat follow the Simplified Technical English guidelines (loosely adapted to my work context)
The problem I’ve been finding is that you can do this but within a few messages, the instructions in the skill will be ignored.
Absolutely infuriating if you’re using Claude in an environment where you can’t run hooks.
Exactly. Sad to see them falling behind on this because it's exactly why I chose to use Claude initially.
They did release an Opus 5 prompting guide saying you need to explicitly prompt it to be concise or it will be very verbose. YMMV but it got better for me to some extent.
https://platform.claude.com/docs/en/build-with-claude/prompt...
And where would we put this? I don’t want to write that out every prompt. CLAUDE.md is a joke, it has little to no effect.
Basically, I’ve gone from supporting them to hoping someone else wipes the floor with them.
Funny, I'm the same. And if find Sol way more pleasant to work with, not to mention way faster. And Sol's compacting is superior, I haven't yet run into it forgetting something crucial from the pre-compact conversation, meanwhile Fable does that notoriously.
When they eventually make Fable available to cheapest plan, I'll downgrade. It's worth keeping for reviewing the code and the UI tasks, but nothing else.
[flagged]
I think anthropic is very far up their own ass and it shows up in the model output
This.
Sometimes a cigar is just a cigar.
I canceled my personal Max 20x subscription because since the 5 series models I simply cannot understand what the LLM is saying without a lot of reading and re-reading, and no amount of CLAUDE.md exhortations to speak plainly seemed to fix it. I don’t have the energy to spend twice as long to understand its plans, and pay Anthropic prices for the privilege. GPT seems not to have been infected by this yet, whatever it is, and Grok is quite refreshing for how normally it speaks.
I wonder if everyone at Anthropic talks like this.
If it’s watermarking, lol, good luck with that, it’s enough negative value to make me switch providers and I’m in a position to make this decision at a company level as well (we spend millions a month on Anthropic).
They need to fix it.
N=2 anecdata but just this week we were discussing setting up a couple of seats with OpenAI as a trial for switching. There are other advantages too, such as being able to bring your own harness including Ai-integrated editors / ACP clients such as Jetbrains, VS Code, and Zed. I think OpenAI and Altman are a clear step more evil than Anthropic and Amodei so I really hate to say it, but with the degradation in model output interpretability, all of the cleverness and power of the Claude Code harness hasn't been enough to offset a genuine falloff in productivity for anything other than total hands-off automation.
That said, the duo of Opus 5 and Sonnet 5 do a fantastic job at fully automated work, and Claude Code still stands head and shoulders above the rest.
I didn't like to use GPT for agentic coding, review yes, but with Opus 5, well I really can't stand anything of that model. I feel that sol xhigh is even better than fable.
Yeah OAI really nailed the communication style with GPT. It also seems just way more token efficient and faster compared to cc. Myself and all my friends have cancelled our $200 Anthropic subs. I'm using a $20 personal plan and even that is enough for my usage so far.
Also using Codex or Pi makes you realise how slow and clunky the cc harness is. Even the desktop app is more responsive and has better UX.
Funny how quickly the tides change.
> Funny how quickly the tides change.
This is something that annoys me working in companies over the years. It’s that you can't just suggest "calm down, chasing the latest thing will not make you faster and is a huge distraction to actual work". Whether it's dot-com tech 20 years ago, latest JS framework 10 years ago, now it's the AI thing of the day. Being calm is interpreted as anti-whatever.
This is 100% my experience.
I think it's a deliberate steganography choice. You can spot Claude vocabulary a mile away, which maybe means you can spot distillations a mile away.
But I agree, the GPT models are so much simpler to work with, they have so much less personality and fewer quirks. They also are a little less aggressive about triple checking every little assumption immediately in a stack of 30 tool calls (but I haven't used 5.6 Sol yet so maybe that's not true anymore).
> which maybe means you can spot distillations a mile away.
I doubt this is the reason. The fact that Chinese labs are all distilling Claude/GPT/etc isn't exactly a well kept secret, they don't even bother removing the name "Claude" from the training data, so the models randomly refer to themselves as "Claude" all the time.
I think it's far more likely to be a side effect of how much synthetic data is being fed back into the models to make them better at coding. The degradation of Claude's prose has been gradual but steady ever since they shifted towards focusing only on code with Opus 4.5.