The Easy Part of Coding Got Automated. The Judgment Part Didn't.

I Thought Coding Was the Hard Part. Then AI Started Writing It For Me. Asked Claude to write me a Lambda function. Done in 10 seconds. Asked it which IAM permissions it actually needed. Silence. So I read the generated code myself. It had asked for s3: , dynamodb: , and a wildcard resource. All Claude did was give the function room to work, not room to fail safely, but I'm the one who has to decide the difference. Turns out the easy part got automated. The judgment part didn't. I spent longer scoping the IAM policy down to specific actions and resources than the AI spent writing the whole function. That's the actual shift nobody's meme-ing about: the code got cheap, review got expensive. Writing was never the bottleneck for anyone who was good at their job. Knowing what "good" looks like still is. What's one thing AI still can't decide for you?