A support lead asks Copilot to draft replies that are empathetic, factually consistent, and moderately creative without sounding robotic. Agents complain outputs vary wildly in tone and invent policies. Which prompt instruction defining response information on sentiment, coherence, and creativity best produces reliable outputs?

Specifying the desired sentiment, coherence rules, and creativity bounds with examples and an output format gives Copilot concrete guidance across all three dimensions. You can ask for empathetic sentiment, define coherence as factual consistency with trusted policy sources, and set moderate creativity bounds so replies sound natural without inventing rules. Examples and a format also anchor tone and structure, reducing variation between agents’ drafts. The instruction that calls for highly inventive creativity fails because it encourages novelty beyond the requested moderate bounds, increasing the risk of unsupported policy claims and uneven tone. The instruction that defines coherence as stylistic flow fails because fluent, well-organized wording can still be factually wrong; it does not adequately guard against invented policies. The instruction that omits examples and a template fails because broad targets alone leave agents and the model room to interpret tone and structure differently, preserving the reported variability. An exam caveat is that naming sentiment, coherence, and creativity is not enough: check what each dimension means and whether its degree matches the task. A practical check is to test the prompt on angry, edge-case, and policy-exception tickets, then assess empathy, policy accuracy, and adherence to the requested format before rollout.