[perrygeo]

Great point, this maps to my experience. Almost every "fix this existing code" problem posed to LLMs is a failure. The approach of simply spitting out the next probable token doesn't work if any complexity or design aesthetics is involved. We wouldn't trust a human that approached problems by stream of consciousness either.

But vibe coding does work reasonably well for solo greenfield/demo projects. When there is no external complexity, no inferred requirements, no consequences to getting it wrong, nothing to integrate with, nobody to collaborate with ... just you and a blank editor - vibe coding works.

While it can be nice living in a little self-contained fantasy world, on existing projects you'll find absolutely none of the same conditions. And a whole slew of other engineering challenges that aren't solved by throwing generated code at it. As the value of code trends to zero, the value of those other engineering skills increases.