Changing one word in a prompt can reduce large language model inference energy by up to 70%, but no prior work explains why. We introduce GreenSwap to find out. We run every prompt under three conditions: the original verb (C1), a swapped low-energy verb (C2), and the original verb with an explicit instruction to limit the response length (C3). Across three open-weight models, three task types, and 21,600 inferences, the short-answer instruction saves more energy and preserves higher quality than verb swapping in 7 of 9 settings. A token–energy correlation of r ≥ 0.998 shows that inference energy depends only on output length and not on which verb was used. Verb swapping saves energy only because it sometimes makes the model write less, and on long-form generation it backfires by producing a completely different and lower-quality response. Scoped to English instruction-tuned 7–8B models, our results show that telling the model how long to respond is simpler, more reliable, and more energy-efficient than curating verb lists.