AI Coding Weekly

Pocock's prompt of the day is "how do we make this change fully verifiable?"

One question, no numbers, and an agent that goes off and clicks through the frontend by itself.

Matt Pocock's prompt of the day is a single question he puts to the agent. "How do we make this change fully verifiable?"

Matt Pocock
@mattpocockuk
X
Makes the agent bend over backwards to verify all its changes.
Nov 5, 2026 · View on X

He says it makes the agent bend over backwards to verify all its changes, and describes it signing up for real services and clicking through the frontend. His summary is that quality goes through the roof.

What is actually different about it

The wording is doing the work. Asking an agent to test a change gets you the kind of test the agent already knows how to write, usually against the code it just wrote. Asking how the change can be made verifiable makes the agent answer a design question first. It has to say what evidence would prove the change works, and then go get that evidence. That is where the account signup and the frontend clicking come from, because for some changes nothing short of the real path counts as proof.

It also moves the burden. In the usual loop you are the one who decides whether the diff is trustworthy, reading it or running it yourself. Pocock's framing hands that job back to the agent as part of the task, which is a different request than asking for coverage.

How much to read into it

This is one engineer reporting on his own work. There are no numbers, no codebase named, no before and after, and no indication of which model or harness he was using. Quality going through the roof is his impression, not a measurement, and the behaviours he lists are the kind that vary a lot with what tools the agent happens to have available. An agent with no browser and no network cannot click through anything no matter how you phrase the request.

What makes it worth a try anyway is that the cost is a sentence. If the agent comes back with a plan for proving the change rather than a plan for making it, you have learned something about the change before any code moves.

Have the good opinion, not the loud one

One email a day. Under three minutes. What working engineers found out about coding with AI.

Get the next one by email

Coding with AI, read daily so you do not have to. The experiments, the receipts and the arguments from engineers shipping real software. Not a changelog.