Large language models still fall short of human clinicians’ abilities. The solution appears to be a lot more human prompting.