Post

"The code was turbo-ass" might in itself be the best code review ever. I agree, I can't purely vibe code. I don't find it fun, and it doesn't really provide good results on anything non-trivial. Good as an assistant, not main driver.

It’s great at finding the root cause I’ve problems. At work I’ve found I can feed a Linear ticket into it and more often than not 15 mins later it comes back with where the problem is. It did a lot of super dumb shit for me in this project.

Sorting topologically is an excellent idea here. Seems like a painfully obvious improvement in retrospect! As the best ideas are. How does it handle things that are not part of the module graph? Eg a random md file

It currently doesn’t, but one of the things I’d like to do is to group files into subgraphs if it finds multiple graphs without a connection.

I think you could do even better than topological sorting if you had enough static information and calculate all of the paths through the modified code, so you could review logical flows individually. It has potential to grow explosively but that could be a signal the PR is too complex.

Vibe coding pros: - I was able to get a lot of chores done around the house while working on this. - Quite nice to give high level tasks and feedback then come back a few minutes later to see how it’s doing. - I don’t think I’d have been able to make the time to do this otherwise.

Cons: - If you let it do what it wants without reading the code it will produce a nightmare of cthulean proportions. - Very much not an interactive process. Everything takes minutes. Not long enough to focus on anything else but not short enough to feel good watching it happen.

Many times it got stuck on a specific bug and I tried half a dozen times to explain the problem and desired behaviour before having to go and look at the code. Invariably the code was turbo-ass and I spent an hour or so prompting it to be better. Tests and static analysis help a lot.

Are you familiar with plan mode? You activate it by shift+tab’ing in your Claude code session. I’d advise doing this always. You give instructions, then you carefully review the plan and make changes there, which means you fix half the issues before it even writes any code.

The plan can be made way quicker than the changes so feedback loop is much more fun, you don’t have to wait minutes to find out it’s done something stupid. Also were you on Sonnet 4 or Opus? Opus is noticeably better, though it is 10x as expensive.

I’m aware of plan mode but haven’t tried it yet. It’s on my list! I’m also mostly using Sonnet 4. I’ve heard folks say “Opus for planning, Sonnet for execution.”