I Gave AI a Spec for a PHP Framework. Here's What Shocked Me.

I ran an experiment that genuinely surprised me.

I took a detailed specification for a PHP framework I'd normally scope at 5 engineers + 2 years of work. Gave it to Claude (with full constraints, edge cases, the whole thing). Two days later: working code.

Not polished. Not production-ready. But functionally there—handling async patterns, error states, and architectural decisions that felt like they required human judgment.

Here's what surprised me most: it didn't just generate boilerplate. The framework had opinions. Dependency injection. Middleware chains. Testability built in. Things you'd expect from a team that'd shipped similar systems before.

Where it fell apart? Documentation. Type hints in weird places. And—critically—it couldn't reason about scaling decisions I hadn't explicitly stated. It optimized locally, not systemically.

So here's what I'm genuinely curious about (and I think this matters for anyone building with AI):

What's the most impressive output you've gotten from an AI tool? And where did it completely miss the mark?

I'm asking because I think the gap between "this is magic" and "this is a glorified autocomplete" is where the real insight lives—and I suspect your experience shapes how you're building.