Good article. We really need more nuance in the discussion of using LLMs for SE. Not all tasks are the same; we shouldn't expect to use the same process across them!

In my experience, the more constrained the LLM is, either by existing patterns in the code base or by prior art in general, the better the LLM is. When I've asked it do architectural or design work, it has been much less successful. That's one dimension on which LLM capabilities vary. There are certainly more!

People should really start making a distinction between vibe coding without looking at the actual files, on one hand, and using LLMs as glorified autocomplete.

Both can be "100% code written by AI" but there's a world of difference between the two.