When using Claude Code, I had a similar experience to what Marc Schmidt (author of ML/Typescript tools) recently described.
When Claude Code launched, the productivity boost was noticeable. I used it for somewhat complex work. It was impressive and quick.
But as I worked with it a pattern emerged: quality degradation, quick-win fixes/workarounds instead of actual fixes, some actually good code got removed and Claude reframed the new buggy code as pre-existing.
This continued until the point that babysitting Claude exceeded the value produced.
Marc’s description resonates too because this failure mode was consistent.
Claude Code seems optimized for local success rather than global correctness, and seems tuned to say "done" rather than to preserve good code when changing around it.
Codex 5.2 seems to be better at handling large context, and has more respect for existing structure.
Not perfect! But did not require as much babysitting from me.
