I wish they'd found a larger sample, but this is still interesting. The fact that much of the time is wasted on prompting makes sense. One important thing not mentioned however is that much of the market for vibe-coding tools is non-developers, i.e. people who don't have the knowledge to actually do it themselves.
post
Using ai to code is like knowing how to properly google pre ai. If your a shit coder ai won't make you more productive.
The bigger issue I see is that it also results in less experienced coders creating code that might work, but that they don't understand.
And in reality it doesn't work, or only works in very specific scenarios and thus fails with no one who wrote it around to understand why it might have failed.
If you already know what you're doing, AI generating code is redundant. If you don't know what you're doing, it might work for you, up to the point you're spending all your time debugging hallucinatory code.
Yup, yesterday I decided it was finally time to write unit/feature tests on a project that had gotten way to far aling with 0 tests. Just over an hour of babysitting Codex and now I've got 800 tests covering 16 models and 60+ endpoints (8000 lines of code) in the exact code style I wanted based on my AGENTS. Honestly, reading through the tests, I wouldn't know that I didn't write them!
Now, trying to use AI to solve a novel problem, or even a not so novel one with any logical complexity, and hours can easily be wasted trying to guide it, correct it's mistakes, and then eventually roll it all back and do it yourself because your smart enough, just being lazy! 😬
top 3 comments