This is quite insightful. Often in my projects, some central nouns come up over and over again in discussions about architecture, until they become embedded in the actual software and the interface.

Since I started using AI tools to assist, I've found a lot of both utility and frustration revolves around my use of these nouns in prompts (in the context of, e.g. "during the 'quickpay' confirmation phase..."). When the bot settles into understanding these nouns, it seems to get a better handle on the architecture as a whole. When it suddenly forgets them and has to go figure out what they mean by scanning the code base, I know it's about to do something staggeringly redundant and stupid.