> use Claude Code for months
> Claude makes wrong assumptions every session
> overwrites code I didn't ask to touch
> adds 500 lines when 50 would do
> find the CLAUDE.md file on GitHub trending
> 82,000 stars. one file.
> paste it into my project
> first task
> wait. it asked before assuming?
> only touched what I asked?
> 50 lines instead of 500?
> pause. read the 4 principles.
> think about every broken diff I reviewed
> every rewrite I didn't ask for
> every session that ran with wrong assumptions
> it didn't have to be like this
> one file. everything changes.
> skill issue discovered
Codex and Claude don’t converge. Each has strengths and weaknesses. I A/B test constantly (using https://t.co/Pftik3Sha1) here’s what I’ve found:
Codex is *substantially* better at long-running tasks. It stays on task and does what you ask it, almost regardless of the complexity of the task, provided it’s well described. If you add more context while it’s working, it doesn’t lose its objective. It’s incredible. But if you use it for any form of UI, whether a high-level design or a small tweak like adjusting a dropdown, it has zero taste and very poor spatial understanding.
Claude cannot do this. It cannot stay on a long task. When folks tell me they can do anything Codex can do with Opus, I immediately know what their programming pattern looks like—it works, it’s just much less parallel. It’s how I coded back in the Sonnet days. Claude cannot receive new context while it’s working without going off track, and it is almost useless without plan mode. In a large multi-step plan, it often leaves out entire portions of that plan. When tests fail, it convinces itself, and you, that those failures are not related, even when they obviously are. It’s not a good model for serious work. *Except* UI and design—it is tasteful, it has excellent intuition about spacing, its mastery of CSS is significantly better than Codex, and it’s able to bridge canvas design with HTML design, which is totally unthinkable for Codex. Even one-shotting a marketing site will likely produce acceptable results. It’s really impressive.
It’s hard to do serious work without both models right now. If I had to pick one, I’d still pick Codex, because you can brute-force Codex to do okay UI, but you cannot do massively parallel Claude Code.
Run your own A/B when doing your daily work: https://t.co/Pftik3Sha1