Exploiting Primacy Effect to Improve Large Language Models