I use AI to correct AI :D Yea I know, sounds ridiculous, but hear me out.

Claude Web works differently than Claude Code. I use Claude Web/Chat for briefs, general overview of the project outside of code/app, for planning and handing off the work to Code, that includes prompts. Code is tasked to output a detailed view of what it found, run measurements and tests (designed both by me and Chat) and report findings back. That is fed back to Chat and I get a rough summary. Errors are categorized by their blast-radius and so on.

Eventually I do a deep dive of what changed, how well it's handling, how much it's struggling, token usage, time spent per review/test and so on. (btw, Code is tasked to run an independent review via contextless agent on the codebase, that easily costs me another 20-30 mins and ~200K tokens on my current project, each run).

It still makes architectural issues, tries bandaiding and patching instead of properly designing, overvaluing/devaluing certain cases and misallocating where most of the focus should be, but that's relatively easier to spot and propose a fix for than doing a line by line code review.

Codebase quality is ensured through mapping the flow and the output space and having a hard-pinned tests Code can't modify to green its answers, + the independent reviewer I mentioned above.

It's not as fun as going in by hand on deep technical problems, sure. But AI usage is less about fun and more about output speed, so use it for what it is for ig? :D