[bayareapsycho]

I don't really feel this kind of friction. In deep coding sessions it usually follows this loop of

1. cook up design

2. coding

3. compiling+running it

4. view the logs, figure out what broke

5. back to 2

AI just makes 2 and 4 happen faster. Which frankly just makes things easier. I don't have to worry about how "modern" my C++ is because it already does it for me. And for debugging, it just does the cmd+clicking through the codebase for me.

So it really just makes it less fatiguing, especially for moving around large bodies of code and renaming stuff. I can spend more time on the essential problem.

I think having multiple of these loops running at once (or having an agent just iterating on its own) is kind of dumb tbh and I don't use them that way. I think having 100 agents running at once or whatever the fuck these people are saying is bullshit. Just using it to speed up 2 and 4 is good enough for me (and also using to explain what the code is doing for building my mental model).

Usually step 3 takes a long time as well, so if claude alone vs claude+me is less accurate, that gets amplified. Another reason why I don't like to let it run all by itself