Pocock trims lauren's prompt to five words
A misunderstanding check you can type in five words, borrowed from a longer version that lauren says is one of her most used prompts.
Matt Pocock says he has started asking coding agents to restate his intent before they continue. The prompt is exactly that, five words. "Restate my intent before continuing."
Very useful for when you're dictating out a bunch of stuff and you can only half-remember what you've said.
It is a trim of a longer prompt lauren posted, which she describes as one of her most used recently. Hers asks the model to restate in its own words what it thinks her goals are and what problem she is trying to solve. Pocock says he prefers the shorter form.
restate in your own words what you think my goals are and what the problem i'm trying to solve is
When it earns its keep
Pocock is specific about the case where it pays off. He says it is very useful when you are dictating out a bunch of stuff and can only half remember what you have said. That is the situation where the prompt does real work. A long spoken brief goes in, you lose track of your own framing somewhere in the middle, and the agent quietly picks whichever reading of it is easiest to act on. Asking for a restatement surfaces that reading while it is still cheap to correct.
Neither post claims a benchmark or a measured improvement. This is two practitioners reporting a habit, not an evaluation. What makes it worth copying is the cost. It is a single line typed before the agent starts producing output, and the failure mode it catches, an agent confidently solving the adjacent problem, is one most people reading this have paid for at least once.
The difference between the two versions
The longer version asks for two things, the goals and the problem. Those are not always the same, and separating them is the part that does the work when your input was unstructured. You can want a function to stop throwing while the actual problem is that the data model is wrong. A model that restates both has to commit to an interpretation of each.
Pocock's five words collapse that into one word, intent, and trade some of the specificity for something you will actually type every time. Which matters more depends on how often you use it. A prompt you remember beats a better prompt you skip.
For working engineers, the thing being tested here is not the model's reasoning, it is the handoff. Both versions are read-before-write checks on the brief, and they run before the agent spends tokens on an implementation. If the restatement comes back wrong, you have learned that in one turn instead of after a diff.