I reached a point with Tutorroo where the code being generated was far beyond my ability to comfortably read. My JavaScript is... rusty, at best. Claude was churning out React hooks and state management logic at a pace that felt great when it worked, and terrifying when it didn't.

The problem was clear: I had no way to verify if the implementation was actually solid, or if it was just a house of cards waiting to collapse the next time I asked for a feature tweak. I couldn't manually QA every edge case.

The constraint? I wasn't about to pause the project to go take a three-week React bootcamp. I needed to ship this thing so my kid could actually use it, not become a frontend architect.

The solution was a bit absurd but entirely necessary. Codex, you're in.

I started using one AI to audit the other. If Claude wrote a complex component, I'd hand the file over to Codex with a simple prompt: "Hey Codex, is Claude hallucinating this React hook? Is this actually how you handle state here?"

It became a strange new workflow. I wasn't writing the code, and I wasn't even really reading the code. My primary skill had shifted to being the most annoying QA lead in the room, cross-examining one bot's work with another bot's judgment.

And honestly? It worked. It caught weird prop drilling, it pointed out performance traps, and it kept the codebase from turning into complete spaghetti. It is a weird way to build software, but it keeps the project moving without requiring me to become the expert.