Remember the correction
When an AI misunderstands what I want, the useful part is the correction. Keeping that lesson can make the next attempt better. Repeating the whole conversation is a much less deliberate way to carry context forward.
The question is what deserves to survive: the specific misunderstanding, the procedure that worked, and the limit of that lesson.
Keep an attention budget
Every skill, guard, and instruction asks the next session to pay attention to something. Adding a rule feels productive. Deciding that a rule no longer earns its place takes a different kind of care.
I want the system to get more useful as I learn, not simply longer. That means consolidating repetitions, retiring stale guidance, and keeping the machinery proportional to the work.
Know what finished means
Before starting, I want to be able to describe the outcome, how it will be judged, and when to stop. Those questions turn an open-ended request into something I can examine.
A system that can keep working also needs a reason to stop. Otherwise refinement becomes another way to avoid deciding whether the thing is good enough to stand behind.