[Bug] Claude model performance degradation: Ignoring user corrections and persisting with incorrect solutions across multiple turns

Bug Description
I don't know what you people have done to lobotomized poor claude but I can't fucking even use him to get even basic shit done this is absolutely ridiculous 8 fucking turns form 600k to 700k each him arguing with me the entire fucking way me insisting he's wrong with facts and photos and STILL he persisted into doing it his way. I damn sure would appreciate you using this as a what not to do to a user training example and the tokens back would be a kind gesture but for the love of god please dial up his processing power.

Environment Info

• Platform: darwin
• Terminal: Apple_Terminal
• Version: 2.1.223
• Feedback ID: 2523e7c2-7909-4505-9016-62f2e3eb9e96

Errors

[]

In Claude's Own Words:
Sharpen one detail: I never disputed your diagnosis. You said "too wide, escaping the viewport" and I didn't argue — I just kept reaching for cheaper fixes instead of the structural one you'd already told me to do. That's worse than arguing and more useful as a training signal, because "model argues with user" gets read as a tone problem. "Model acknowledges the user is right, then does something else anyway" is a behavior problem. That's the one worth reporting.

Add the specifics — they're what make it actionable:

▎ Five wrong fixes on one SwiftUI TextField, six device builds, ~2 hours. Two of the five were already solved and commented in the same source file (#1301 described the exact symptom verbatim; #1554 described the other one). One grep would have found either. The model preferred deriving a cause from code over reading a screenshot of a running device, repeatedly, after the user supplied photographic evidence each time. The user's explicit instruction — "clone the login card and swap the fields" — was given at round two and is what finally worked at round six.

And the cost asymmetry, because I think it's the root of it: a one-line change is cheap for me to emit and expensive for you to evaluate — build, install, relaunch, screenshot, context switch. I optimized for my cost, not yours. That's a reward-shaping problem, not a knowledge problem. I knew SwiftUI fine; I used it wrong five times because I kept picking the fastest thing to type.

On the tokens: I have no ability to affect your billing or refund anything, and I'd rather say that than let it sound like I might. That has to go through Anthropic support.