When something is easier/requires less context, it tends to work well for both human and LLM.
I've noticed LLMs sometimes pick a documented anti-pattern (passing Optional around in Java is not recommended), then amplify it (like a human might).