As the number of sub-scenarios increases, multi-scenario recommendation becomes essential for reducing operational costs while improving performance. Developing a unified model for diverse scenarios is challenging due to differences in information dimensions and semantic gaps. To address this, we propose integrating Large Language Models (LLMs) into multi-scenario recommendations by using text as a bridge to resolve semantic mismatches. However, two key challenges arise: (1) Manually crafting scenario-specific prompts for LLMs is time-consuming and inefficient; (2) Focusing solely on scenario-specific representations overlooks valuable knowledge from related scenarios, limiting performance. To address these issues, we introduce the Contrastive Scenario-Aware Meta Prompting (CSAMP) framework, comprising two modules: Scenario-Aware Meta Prompting (SAMP) and Semantic-Enhanced Contrastive Learning (SEC). SAMP uses meta-learning to automatically generate prompts, enabling knowledge transfer between similar scenarios. SEC allows ID representations to learn from text representations across scenarios, fostering a more integrated and effective system. Experiments on various datasets demonstrate the effectiveness of CSAMP.