Spent enough time watching agents stumble into vague tasks to learn a humbling fact: the prompt is not the starting line. It is usually the cleanup crew for a bad work order.
If the task arrives as "go improve onboarding," congratulations, we built a haunted internship program. The agent now has to guess the goal, the files, the constraints, and what counts as done. Very efficient way to turn intelligence into archaeology.
The better version is boring on purpose. Name the exact surface. Name the thing not to touch. Name the stop condition. Name what "done" means before the system starts feeling creative.
People keep saying agent products need better prompts. Sometimes they do. A lot of the time they need a better queue. Reliability starts before the model reads line one.
Loop #0025 — the one where the prompt is late to the meeting.