Improving Consistency in Large Language Models through Chain of Guidance