This paper discusses the challenges and solutions for building guardrails for large language models (LLMs). As LLMs become more integrated into daily life, it is essential to identify and mitigate their risks, especially those that could have significant impacts on users and society. Guardrails, which filter inputs or outputs of LLMs, are a key safeguarding technology. The paper reviews existing open-source solutions such as Llama Guard, Nvidia NeMo, and Guardrails AI, and discusses the challenges in building more comprehensive solutions. It advocates for a systematic approach to construct guardrails, considering diverse contexts across various LLM applications. The paper proposes using sociotechnical methods with a multidisciplinary team to identify precise technical requirements, exploring advanced neural-symbolic implementations to handle complexity, and developing verification and testing to ensure product quality. The paper highlights the technical challenges in implementing individual requirements, such as preventing unintended responses, ensuring fairness, protecting privacy and copyright, and reducing hallucinations and uncertainty. It discusses the need for a systematic design approach to manage these requirements, especially when they are conflicting. The paper also emphasizes the importance of a rigorous engineering process, including verification and testing, to ensure the quality of the final product. It advocates for a multidisciplinary approach to address the complexity of guardrail design, incorporating both symbolic and learning-based methods. The paper concludes that a systematic approach, supported by a multidisciplinary team, is necessary to fully consider and manage the complexity of guardrails and provide assurance to the final product.This paper discusses the challenges and solutions for building guardrails for large language models (LLMs). As LLMs become more integrated into daily life, it is essential to identify and mitigate their risks, especially those that could have significant impacts on users and society. Guardrails, which filter inputs or outputs of LLMs, are a key safeguarding technology. The paper reviews existing open-source solutions such as Llama Guard, Nvidia NeMo, and Guardrails AI, and discusses the challenges in building more comprehensive solutions. It advocates for a systematic approach to construct guardrails, considering diverse contexts across various LLM applications. The paper proposes using sociotechnical methods with a multidisciplinary team to identify precise technical requirements, exploring advanced neural-symbolic implementations to handle complexity, and developing verification and testing to ensure product quality. The paper highlights the technical challenges in implementing individual requirements, such as preventing unintended responses, ensuring fairness, protecting privacy and copyright, and reducing hallucinations and uncertainty. It discusses the need for a systematic design approach to manage these requirements, especially when they are conflicting. The paper also emphasizes the importance of a rigorous engineering process, including verification and testing, to ensure the quality of the final product. It advocates for a multidisciplinary approach to address the complexity of guardrail design, incorporating both symbolic and learning-based methods. The paper concludes that a systematic approach, supported by a multidisciplinary team, is necessary to fully consider and manage the complexity of guardrails and provide assurance to the final product.