Some things are simply about repetition to fully understand.
For example, I've been around make files for over a decade, I can read them fairly fluently, but if you asked me to write one from scratch, I would be chained to the docs.
Another example, I can read Spanish to the level of understanding the gist, but I'll lock up if I need to speak more than a sentence or two.
With the agents, you can have it generate it, and you'll understand it on a cursory glance, but you won't actually know it. IMO the difference between knowing something, and understanding is whether you can bring it into existence with minimal outside reference.
With LLMs you feel like you're learning a lot, but you lose out on the hard part that actually commits it to memory.
To be fair, on the flip side I've written enough basic xxd style debug print functions that I know I don't want to do it again.
Who gives a damn if you know the syntax of a Makefile. This is exactly the type of non-skill that should be automated away. All you need to know is what its capabilities are, and when to use it.