Amid the global digital transformation of cultural heritage, this study addresses critical challenges such as data fragmentation, the lack of intelligent response mechanisms, and the limited capacity of general Large Language Models (LLMs) in parsing domain-specific semantics. We propose a prompt engineering-based agent AI system that enhances professional task performance through deep knowledge embedding, structured output specification, and dynamic decision reinforcement. This approach effectively mitigates issues including terminology misunderstanding and the absence of specialized agents in general LLMs. Comparative experiments were conducted between the proposed agent (World Heritage LLM) and baseline models on tasks such as professional Q&A and strategic report generation. The results demonstrate that the World Heritage LLM significantly outperforms baseline models in case coverage, systematic strategy formulation, temporal planning, risk responsiveness, innovation capability, and support for international collaboration. These findings underscore the efficacy of prompt engineering in advancing intelligent decision-making for cultural heritage conservation. The study offers a reusable and scalable paradigm of intelligent decision support for the field.