Pocock's prompt of the day asks the agent what it wanted to ask
A one line prompt for the moment an agent has been working alone for hours.
Matt Pocock posted a one line prompt of the day for long unattended agent runs. The prompt is "/wait-what questions did you have for me". He suggests it for the case where, in his words, the agent "has gone off on its own for hours and has random questions scattered through its context window".
Great for when the agent has gone off on its own for hours and has random questions scattered through its context window.
That is the whole thing. No benchmark, no before and after, no claim about which model or harness it works best in. It is a habit from someone who posts these daily, not a measured result, and it should be read that way.
Why the questions pile up
An agent working on its own does not block and wait for you. When it hits an ambiguity, a naming choice, a missing requirement, a guess about what you actually meant, it picks something and keeps going. The reasoning around that choice stays in the context window, but it never reaches you. By the time you come back to a finished diff, the decision looks like code rather than a question.
Asking for the questions after the fact flips that. Instead of reading the output and trying to reverse engineer where the agent improvised, you get the list of places it was unsure, in one pass, from the only thing that still remembers them.
For working engineers
The cost is one prompt at the end of a session, which makes it cheap to try on your next long run. The limit is obvious too. The agent only surfaces uncertainty it noticed. The decisions that bite hardest are usually the ones it made confidently and wrongly, and those will not appear on the list.
It lands in the same territory as Pocock's other prompts, which keep circling the gap between what you asked for and what the agent decided you meant. An earlier one worked the front of that gap by making the agent restate your intent before it starts. This one works the back end, after the damage, or the acceptable guesses, are already in the tree.