Pocock runs /grill-me one question at a time on fuzzy specs

Matt Pocock says he still sometimes types "/grill-me about this, one question at a time" at his agent. The trigger, in his words, is when he REALLY has no idea what he is building, when the problem is hugely underspecified and a single question and answer round could send things in massively different directions.

Matt Pocock

@mattpocockuk

X

Usually when I REALLY have no idea what I'm building. When the problem is hugely underspecified and a single Q&A round could send us in massively different directions

The interesting part is the modifier, not the command. A grill-me style prompt normally produces a batch of clarifying questions, you answer them all at once, and the agent then writes a plan off that one exchange. Pocock's version throttles that to a single question per turn. Each answer you give shapes what gets asked next, which is the whole point when you do not yet know what you want.

The escape valve

He also mentioned, for folks asking, that he uses a separate /wayfinder prompt as an escape valve for when a task will obviously take more than one session to plan. He did not say what is inside it. So there are two tiers here. Short fuzzy problems get the one-question drip. Problems that are visibly too big to settle in one sitting get handed to something else entirely instead of being forced through the same interrogation.

Why the pacing matters

For working engineers, the copyable thing is the pacing, not the skill file. An agent that asks you ten questions in one message is asking you to specify the whole shape of the work before it has learned anything from your first answer, and most of those questions were generated from its initial guess about what you meant. If that guess was wrong, every question after it is wrong too, and you are now ratifying a direction you never chose. One question at a time costs more turns and gives you a chance to redirect on each one.

This is one engineer describing a habit, not a measured result. Pocock offers no before and after, no comparison against a batched grilling, and no claim about output quality. Take it as a prompt you can try on your next underspecified ticket rather than evidence that it works better.

It does fit a pattern in how Pocock talks about agent setup, which is that the win usually comes from shaping the conversation rather than from the tooling around it. He has previously argued for building shared vocabulary with the agent through a GLOSSARY.md file. Same instinct here. The agent is going to commit to an interpretation of your request at some point. The question is whether that happens before or after you have told it anything useful.