Advising to homebrew everything with an LLM just sounds insane to me. Cognitive overhead of what? Using well thought out and standardized libraries that have proper documentation? Creating a new engine with an LLM is only going to achieve the opposite of fully understanding your codebase.
Let's say you're making a visual novel. Do you need a fully 3d editor? No, so having Z coordinates on all your objects is a liability: you'll have bugs with Z-fighting sprites or your designer will accidentally place gameobjects at depth and not understand why they're not rendering, because this is a 2d game and they set the rendering layer correctly not thinking about 3d space.
Specify exactly what you need and execute that, using libraries for the low level stuff, and taking responsibility for as many design decisions as you can. It will save you so much frustration. If I could go back and give one piece of advice to myself when I was learning, it would be ditch the engine - they are false promises. One size fits all means it truly fits no one.
You are right that you can often ditch the engine, but you should only do that as an experienced programmer who can assess the scope of the project. AI doesn't really change that IMO.