Let me prompt your whatever language model you're using that selects clips. It currently seems to be biasing towards informative, but the vibe I'm going for is rhetorical. If I can provide a prompt describing the types of clips I'm looking for, I think the recommended clip quality might be higher?