In response to the challenges faced by enterprises in accessing industrial policy information due to lengthy and terminology-dense texts, as well as issues of factual distortion and missing elements in summaries generated by general-purpose large language models, this paper constructs a multi-source text dataset comprising 200 policies from 2020 to 2024, covering 10 categories at national, provincial, and municipal levels. A standardized prompting framework incorporating element coverage templates, factual constraint rules, and output format specifications is designed. Experimental results demonstrate the effectiveness of this prompting framework, providing a deployable solution for government intelligent services and enterprises’ precise policy access.