[johnnyaardvark]

I've heard this take a few times, but I'm not convinced using general language is the new way to write code (beyond small projects).

Inevitably AI will writes things in ways you don't intend. So now you have to prompt it to change and hope it gets it right. Oh, it didn't. Prompt it again and maybe this time will work. Will it get it right this time? And so on.

It's so good at a lot of things, but writing out whole features or apps in my experience seems good at first, but then it turns out to be a time sync of praying it will figure it out on this next prompt.

Maybe it's a skill issue for me, but I've gotten the most efficiency out of having it review code, pair with it on ideas and problems, etc. rather than actually writing the majority of code.

[seanmcdirmid]

Until you've actually done it yourself, it will probably sound like vapor ware. The only question is how much energy are you willing to spend, in terms of actual energy (because you are making more calls to the AI) and yes, setting up your development pipeline with N LLM calls.

It is really like micro-managing a very junior very forgetful dev but they can read really fast (and they mostly remember what they read for a few minutes at least, they actually know more about something than you do if they have a manual about it on hand). Of course, if its just writing the code once, you don't bother with the junior dev and write the code yourself. But if you want long term efficiency, you put the time into your team (and team here is the AI).