Ask HN: How much of your token use is fixing the bugs Claude Code causes?

My experience using Claude Code is that I will detail work I want it to do, then it does it, then I spend quite a bit of time with it trying to figure out why it didn't work. Putting in debugging sometimes feels like half of the code it is writing. At the same time, I'm using more tokens by default with this "better" version of Claude. Of course I don't think this is 'on purpose' but it feels like there should be some discount on tokens being used to fix the bugs being caused by the bugs being introduced. Curious what others' experience are and if you've developed any strategies to address this.