Hacker News new | past | comments | ask | show | jobs | submit
LLMs can write high quality code, much higher quality than even the best engineers, if you prompt it at a low-level. You make the design decisions, design the abstractions, the data flow. The AI just implements logic.

What LLMs can’t currently do is be allowed to run autonomously (agentically) and fulfill feature requests and bug fixes. They will fuck up the code base beyond belief and create spaghetti.

When it comes to discrepancies in how people experience LLM programming, I believe the distinction comes from this. You have to prompt it at the right level of abstraction. If you ask it to do big things, it will do it, and burn down the world in return. But if you carefully design and scaffold and API and write tests, it will implement it flawlessly.