I want a review that explains a concrete problem, because a vague second opinion only adds to my reading queue. Sonnet implements, Opus reviews, and I like separating those responsibilities. Adding another model to the conversation does not, by itself, establish which responsibility it received.
Implementation tries to produce the requested change. During review, I want the reviewer to examine its consequences against the expected behavior. If the second agent is merely told to continue, it may deliver more implementation. Then I have another executor and still lack the reading I wanted. The job needs a name before the model gets assigned.
So the review request is a prompt rule, and it has four parts: the objective of the change, the diff, findings tied to that objective, and older problems marked as older. A question like "what happens when an input is incomplete?" belongs in it. The reviewer should locate the relevant path and explain its conclusion with something I can check. A generic recommendation to improve robustness leaves me doing the work of discovering where a problem exists. Thanks for the essay topic; now I still need the review.
A capable model cannot fill in a missing requirement. If expected behavior exists only in my head, even a capable reviewer needs to infer it or ask. Making the expectation accessible is my responsibility, along with delimiting which consequences belong in this pass.
When the reviewer finds an older problem during the reading, I want it marked as older so I can decide whether it belongs in the current task. Comments also need a destination: accepted findings become a change or an explicit scope decision; rejected findings need an understandable reason. If the findings arrive as a loose list with no decision on each one, the next agent ends up choosing which ones to apply.
I will keep the roles open to different model assignments and judge reviews by the usefulness of their comments. A good response should let me decide what to do next. Presenting the change again in prettier words does not justify another round.