BTW, if it doesn't take you hours to test the failure modes, you're not thinking of enough failure modes.
The time savings in writing it myself has a lot to do with this. Plus I get to understand exactly why each line was written, with comments I wrote, not having to read its comments and determine why it did something and whether changing that will have other ramifications.
If you're doing anything larger than a sample React site, it's worth taking the time to do it yourself.
The main key in steering Claude this month (YMMV), is basically giving tasks that are localized, can be tested out and not too general. Then you kinda connect the dots in your head. Not always, but you can kinda get gist of what works and what doesn’t.