We tested fifty prompts across three models for six weeks, and the results upset the prompt engineering folklore. Clever prompts, the ones with personas and elaborate theater, performed worse than boring structured ones. What actually correlated with quality was structure: clear input format, explicit constraints, and a place for the model to show uncertainty. The single biggest improvement came from adding one line: if the input is insufficient to answer accurately, say so and list what is missing. Hallucinations dropped by more than half because the model stopped filling gaps with confident fiction. Stop roleplaying as a senior consultant with fifteen years of experience. Start writing prompts like contracts. Specify the input, the output format, the constraints, and the escape hatch for insufficient information. Boring wins again.