One prompt drafts alternative marketing taglines and should produce a fresh set each run. Another assigns each support ticket one of six fixed labels and must assign the same label to the same ticket every time. Both run on the same model. How should the generation settings differ, and why?

The principle — a generation setting is chosen from what the task requires, and these two tasks require opposite things from the same knob. The question to ask is not which value is better but whether varied output is a feature or a defect here.

Why the key is correct — Google states that lower temperatures are good for prompts that require a more deterministic or less open-ended response, while higher temperatures can lead to more diverse or creative results, and that a temperature of 0 is deterministic because the highest probability response is always selected. Taglines are the open-ended case, so variety is the product. Labels are the deterministic case: the same ticket must receive the same label, or the classifier's output cannot be compared or trusted between runs.

Why the others are wrong — running both at 0 and rewording the prompt for variety works, but it makes you hand-write what the sampling would give you for free. Raising both treats variety as a general quality improvement, when for the classifier it is the defect. And leaving both untouched assumes sampling only matters for creative work, which is precisely backwards — the classification task is the one where run-to-run drift causes real damage.

Remember this — ask what the same input should do on a second run. If a different answer is a feature, allow randomness; if it is a bug, remove it. The setting follows from that, not from a default.

Sources — Google's Prompt design strategies.