When LLMs struggle: Architecture, context, and hidden complexity

The obvious LLM failures are easy to catch. Syntax errors, broken configs, a pipeline that refuses to run. You see the problem immediately and fix it. Those are not the ones that should worry you. The ones that should worry you are the ones that look completely fine. The code runs. The config is valid. The output looks reasonable. And yet, when someone with more experience takes a look, the problems become obvious immediately. They were just invisible to you. The knowledge mirror There is a pattern I noticed pretty quickly when working on tasks outside my main area. When I work in areas I know well, like Terraform or CI/CD pipelines, I can evaluate the model's output almost automatically. I know what good looks like, I know the common failure patterns, and I catch mistakes fast. The feedback loop is tight. But when I work on something I know less about, that feedback loop breaks. And the problem is that the model does not help you here at all. It does not become more cautious in unfami