When an answer misses, most people delete everything and rewrite the prompt from scratch — throwing away the model's context along with its mistake. Operators do the opposite: they treat the first output as a draft and issue a correction. 'Keep the structure, halve the length, drop point 3.'
Iteration works because the model already holds your task in context; a targeted correction is cheaper and more precise than a cold restart. Restart only when the approach is wrong, not the execution.
In the Lab, each run is a fresh start — which makes it the perfect place to practice the other half: tightening the prompt itself between runs. Run the seed, spot the weakest part of each answer, fold a fix into the prompt, run again.