If you’ve used ChatGPT, Claude, or any of the current LLMs to generate code, you’ll know the drill: they almost get it right… and then don’t.
For anything beyond the most basic requests, they often miss the mark — and suddenly you’re deep in a debugging rabbit hole, politely wrestling with a very confident but very wrong assistant that keeps firing new versions at you like a spaghetti cannon.
The real issue?
These models try to boil the ocean instead of solving problems in small, logical steps. They skip the fundamentals we learned in the early days of programming — breaking things down, testing as you go, building with reusable, modular blocks.
What they need is a crash course in structured thinking and a reminder that general-purpose subroutines still have, well, a purpose.
Until LLMs get better at thinking like engineers rather than overly eager interns, we’ll have to hold their hands — or better yet, teach them to break things down properly in the first place.
Anyone else seeing the same pattern?
Would love to hear how you’re steering your AI to code smarter — or at least wrong less enthusiastically.
Let us know what you think ….