LLMs have shit memory and abstraction capabilities and just adding md files and more context is like trying to make a alzheimer patient learn to play the piano. Yes it kinda works to some degree to always repeat everything, but it's not true learning and just a bandaid.
It's an implementation detail though. For example when you do multiple tasks in the same repo with the same rules, your context stays cached and you're not "repeating everything" anymore. Does that make a difference?
I mean that's certainly one opinion, but I mean, as you say, it does work.