It's own code output will generally work for things that have been done before, but quickly fall apart when you are trying to create something new. You'll need to write a lot of pseudo code to make that work, which is actually excellent training as well.

It really does help if you instruct the language model to be honest and to not do it's very best to please you.

I have not found a solution to the repeating incorrect answer loop yet, although some environments offer a "banned response" list, that can mitigate the situation a bit.

As to the OPs question, yes, definitely use it! But more importantly, learn how to use it. Your approach sounds reasonable and similar to mine. Companies that ban the use of LLMs or developers that scoff at it's benefits are delusional.

Something that is easier to use might come eventually, but in the short term future I expect we will see a clear separation between developers with strong grammatical skills who have learned how to prompt and those without the ability or the experience.