You can do a task pretty well if you nudge the AI, have it write an exact explanation about every part of the architecture, code and data flow it's working with and throw relevant files into context, and correct anything that's wrong before you send it to do the task. You still have to review, but I didn't have to correct much in my experience.

But that burns like 20$ of tokens per task, at current prices that are way below the costs AI companies are paying.

While it does help me, especially with parts of the codebase I'm not familliar with, it's not sustainable, and it's actively and very quickly robbing me of my skills and knowledge. It's really a bad idea to use it, in two years time you'll be royally fucked once they raise prices to recover the trillions they are loosing right now.

So, however tempting, I simply don't use it. I won't throw away years of college and experience just to do a task a little bit faster today.