I think it's because LLMs in single shot scenarios are always liars, fabulists, or both.

The only reason (and only way) it works with coding is because you can hook it up to tools to run static analysis on its bullshit and it can guess enough derivatives in a loop to eventually build something that kinda sorta works.

This doesn't apply for many fields outside of writing software. There just aren't many fields where it applies to have a machine produce permutations and then automatically run validation on the result.

Imagine if you put it in charge of manufacturing or something, you'd have so much wasted raw material before you got to an acceptable result that you'd go out of business.