Which prompt engineering technique should be used to improve a large language model's response quality for complex problem-solving tasks that require detailed, step-by-step reasoning?