Opus 5 is actually almost rage-inducing to use.

I've read the latest recommendations from Anthropic, I've tried all kinds of global claude.md file changes, but I just cannot make Opus 5 behave in a way that's pleasant to work with.

I'm not talking about if it's "nice" or "pleasant" in its phrasing, I'm talking about the following:

- Responses are way too verbose and buzzwordy and hard to follow. I legit don't read 90% of the output anymore, that's how bad it is. No matter what I put in my claude.md when it comes to communication style, after it did any type of meaningful work, it always reverts back to its extremely verbose, over-explained, buzzword heavy mess. Huge sentences & paragraphs which contain multiple concepts at once, using buzzwords constantly that serve almost no purpose except making it harder to parse, and just overall doesn't stop yapping.

- Simple tasks turn into these huge "projects" no matter how small something is.

- On the flip side, it just doesn't actually finish things properly. It made a change to a file, which invalidated something else in literally the same file. It then raises to me as part of its insanely wordy response that "I left x unchanged, but it is now no longer valid, let me know if you want it changed as well". It causes so much mental effort just to make sure it doesn't create rot in the codebase. The amount of times the response includes something to the extent of: "Worth knowing: my new change introduced issue x, let me know if you'd like that resolved" is just ridiculous. I feel like I need to prompt it 5 times just to actually finish the work properly.

Don't get me wrong, the model is still useful, it very clearly is still strong at the underlying tasks, but it is just a disaster to work with. Fable doesn't seem to have this issue, or at least to a much smaller extent, but even on the 200$ plan the fable weekly limits are not really enough to use it for most tasks.

Somehow, it has become a socially inept genius, that is both incredibly lazy AND way too overzealous at the same time.

Any tips to make this better are very welcomed, but really I just wanted to vent & provide some feedback because a model like this is just not something I can keep using long-term, I have to look into alternatives for the enjoyment of my work.