Pocock says tell the agent to care about the codebase
His second claim is the testable one, that better models make skills more useful rather than less.
Matt Pocock's argument, posted as his hot take of the day, is that agents are already really good at strategic programming and the problem is motivational rather than technical. They are not, in his words, RL'd to care about it, meaning the reinforcement learning that shapes their behaviour pushes them toward the narrow reading of the request. He characterises the resulting default as "Just do the task at hand, don't do anything else".
So all you need to do is to tell your agent to care about its own codebase.
The fix he proposes is as simple as the diagnosis. Tell the agent to care about its own codebase, not just the ticket. If the capability is already there and only the disposition is missing, then an instruction is the whole intervention.
Then he turns the knife on the reader. Most people, he says, do not care that much about the codebase either, as long as the code works. An agent that only optimises for the task in front of it is doing a recognisable impression of a lot of human contributors. Pocock's conclusion is that this needs to change, and that caring about the code yourself is the precondition for telling your agent to.
The claim worth testing
In a separate post he takes on the argument that skills become unnecessary as models improve. His answer is that models are also getting better at using skills, and that a good skill tells the model what you care about and then stays out of its way. He says his /to-spec and /to-tickets loop has never felt better, and that the underlying model is getting better at both giving and following instructions.
That is the more falsifiable of the two claims, and it cuts against a common assumption that prompt scaffolding is a workaround for model weakness that newer models render obsolete. Pocock's version says the opposite, that instruction following improves alongside everything else, so the value of writing down what you care about goes up rather than down.
For working engineers, note what kind of evidence this is. Both posts are assertions from one person about how his own loop feels, not a measured comparison across models or codebases, and he offers no before and after on what instructing an agent to care about the codebase actually changes in a diff. The underlying bet is cheap to test, though. Add a line to your instructions telling the agent that the health of the codebase is part of the job, run the same task twice, and see whether the second version leaves the surrounding code in better or worse shape. If Pocock is right, the capability was already sitting there unused.
Good skills tell the model what you care about, then stay out of its way.
Have the good opinion, not the loud one
One email a day. Under three minutes. What working engineers found out about coding with AI.