The (Claude Code) Struggle is Real.

friendship meme GPT5 and Claude.
ben taking pics dangerously on the road
ArticleVibe Coding

My little sandbox project is not so little any more as more and more features gets built, and each feature getting increasingly complex. Here's what I noticed: Claude Code is struggling.

While I did enjoy our moments of debugging together which sometimes feel wholesome like as if I'm working with an extremely polite colleague whom I could share all my (app's) issues, to feeling like I'm talking to IT support, telling them what debug messages I see on chrome dev tools. But I feel like it has reached the limit today, at least when it comes to debugging.

I was working (ok not I but AI) on debugging an issue with Mr. Claude and it just kept repeating that it resolved the issue, confidently, but the issue kept persisting. Then it did something disappointing: it hardcoded a certain filter for a particular object I flagged as missing. I proceeded to let it know politely (like how I'd coach a co-worker) that it's not the right way to resolve issues.

But deep down within me, my internal eval for Claude Code took a dip.

Thankfully we managed to resolve the issue after a few more turns. We moved on to iterate on another feature involving complex data grid logic and it struggled again. This time I took a few screenshots,

It went to read and edit a bunch of files thereafter...then:

But the fix didn't work. We went a couple turns more to try and resolve it but nada. Thankfully I've got more pocket co-workers in the toolbox and I fired up VS Code with Augment Code (woo! 7 days free trial!) , and gave the exact same prompt regarding the issue to GPT-5.

And it basically one-shot it.

I went back to Claude and asked it to review my fix (couldn't bear to tell him it was GPT-5's fix) and the replies were once again like an exemplary co-worker who was humble enough to know where he went wrong:

Oh such precious humility! (I can totally get why people fall in love with AI)

With all that said, I think Claude's still pretty good at implementing new features as long as my prompts are well engineered. It's also much much faster than GPT-5 (with less smarts as the trade-off.)

Perhaps moving forward I'll debug with Augment and GPT-5 which hopefully their lowest price tier would suffice.