Depends on the language maybe.

I have had considerable trouble getting C99 code from any LLM that adheres to some basic safety conventions present in all my code. The biggest issue is generated functions have multiple return paths making it more difficult to visually spot cleanup errors.

My usual code has a single cleanup label that frees all resources and, if error is still set to true, cleans up the return value as well. Even when tired this convention makes it dead simple for me to spot errors.

So I've been writing most of my C projects by hand.

SQL otoh, LLMs are great.

True, they're a great tool and shine in some areas. I especially use LLMs to work as a search engine for me before I start a new project; I use it to critique my implementation plan, I use it to find things that already exist to avoid reinventing the wheel, and I use it as a reviewer while building out my spec/design documents. Both of which I first write by hand then use it to format, give me "ideas" which I cherry pick depending on the scope of the project. YMMV.

Also in the same vein as SQL, I've seen Gemini is pretty good at generating formulae in Google Sheets; I used it to create an optimal player rota for a 24 hr online endurance race me and my friends were participating in using parameters like player_strength, the time slots they were fine with taking up, in order to min-max fatigue and consistency.

Pretty fun.