In some ways, refusing to use LLM's to assist with coding is like that funny commercial where they show a guy trimming grass one blade at a time using a nail clipper. Using LLM's responsibly is a massive force multiplier. You still have to know what you are doing or the result could be in a range between not maintainable to disastrous.

In my experiments with LLM's one of the most obvious issues is project structure and organization, particularly in the context of future objectives. It is way too easy to have an LLM go in circles when analyzing ideas for project structures. I have seen them propose layouts that, while workable, are far from ideal.

The same applies to code. It can write usable code faster than a human. However, and, again, in my experience, you are constantly refactoring the results.

In all, this is a new tool and, yes, it is useful. It's amazing to me considering the idea that the whole thing is based on guessing the next token. It doesn't really understand anything, yet, you'd be hard-pressed to prove this to those who don't know how we got here. I've had a few of those conversations with non-coders using ChatGPT for a range of tasks. It's magic.