Large language models (LLMs) often produce inconsistent responses, which complicates their integration into workflows that require specifically formatted outputs, such as queries. To address this, the concept of a "prompt sandwich" is introduced, which structures prompts into three layers: system prompt, context, and user prompt. This method guides the conversation's direction, provides necessary context, and formulates specific questions or requests to enhance consistency and quality. The Elastic AI Assistant exemplifies this technique by integrating data like security alerts and pseudonymizing sensitive information to maintain privacy. Users can customize prompts and manage data anonymization settings, ensuring only selected fields are shared with LLMs. This approach not only improves the reliability of responses but also emphasizes the importance of data privacy when utilizing generative AI.