When users employ LLMs to find suitable ancient Chinese poems or generate prompts for text-to-image models, LLMs often misunderstand poetic elements or metaphors, which leads to false or inaccurate generated content. In this paper, we propose PCOT (Poetry Chain-of-Thought), a method that improves the accuracy of poem content adaptation and poetry image generation by integrating poetry database and verification mechanism to enhance Chain-of-Thought (CoT) reasoning of LLMs. Through these techniques, PCOT is able to better analyze key semantic elements, emotions and metaphors, which then enables more accurate retrieval of poems that align with user intent, and generation of high-quality poetry-to-image prompts. Evaluations demonstrate that our approach outperforms pure LLM in higher accuracy and semantic consistency on tasks of poetry content adaptation and poetry image prompt generation.