Using "underdrawings" for accurate text and numbers | Hacker News

A Hacker News post says the field is getting better at identifying what LLMs do well and where they struggle.

It calls for a clearer taxonomy that matches LLMs to specific subtasks, such as generating functions or SQL, rather than handing them full end-to-end work.

The key idea is to keep humans in charge of structure while letting the model fill in details.

This matters because it helps teams use LLMs more effectively and avoid overtrusting them on tasks that need tighter oversight.