LLMs aren't good when you drift out of the training distribution. You want to be hitting the meat in the middle and leveraging the LLM to blast through it quickly.

That means LLMs are great for scaffolding, prototypes, the v0.1 of new code especially when it's very ordinary logic but using a language or library you're not 100% up to speed on.

One project I was on recently was translation: converting a JS library into Kotlin. In-editor AI code completion made this really quick: I pasted a snippet of JS for translation in a comment, and the AI completed the Kotlin version. It was frequently not quite right, but it was way faster than without. In particular, when there was repeated blocks of code for different cases that different only slightly, once I got the first block correct, the LLM picked up on the pattern in-context and applied it correctly for the remaining blocks. Even when it's wrong, if it has an opportunity to learn locally, it can do so.