Prescribe the steps (left): a meticulous motion SOP — grasp here, lift 30°, rotate, place. It does push the model to a level. But humans execute with drift, and after a long day on one task they collapse to the single most effort-saving path. In learning terms, one mode dominates the distribution, coverage craters, and real-robot performance hits a biased ceiling — efficiency suffers too. Prescribe the result (right): fix only the final test — succeed across many scenarios — and let operators invent how. The distribution stays broad, generalization holds, the ceiling lifts. It's the same reason you don't hand employees a rigid prompt; you hand them the API and hard tasks, and judge the outcome.