As interest in applying artificial intelligence (AI) to safety engineering grows, practical questions remain about its reliability and rigour. This research explores the use of advanced prompt engineering techniques to automate hazard and operability (HAZOP) analysis to refine the output of a large language model (LLM). Focusing on a deployable portable water production system as a case study, we investigate whether LLMs can support the early stages of hazard identification, augmenting the HAZOP process. Our approach integrates multi-step prompting, chain of thought, atom of thought, and few-shot examples to support reasoning and improve the quality of generated outputs. The results are subjected to system expert review to evaluate coverage, accuracy, and usefulness. This study provides an important step in assessing how LLMs can augment traditional safety workflows. We share key lessons from the experiment, including prompt design strategies, system limitations, and practical insights into deploying LLMs in safety assessment. It offers safety professionals a grounded view of where current AI tools can assist in the safety assessment process and where caution is still required.