the annotation cycle in plan.md is the part that actually makes this work imo. it's not just that you're planning, it's that you can inject domain constraints that the model can't infer from the codebase alone -- stuff like "don't use X pattern here because of Y deployment constraint" or "this service has a 500ms timeout that isn't in any config file". that knowledge transfer happens naturally in code review when a human writes the code, but LLMs skip it by default.