AI tests lie. How do you actually verify AI code?

As anon-dev building with AI agents green checkmarks used to be my green light to ship.

Big mistake. An AI testing its own code is just agreeing with itself.

Now before accepting any fix we intentionally break the code. If the test still passes it’s useless. That simple trick caught so many false greens!

What’s the #1 simple check that saved you from shipping a broken AI bug? Drop your tips below! 👇