Large Language Models (LLMs) have emerged as a promising tool for automating CI/CD debugging. However, LLM-based failure diagnosis in GitHub Actions (GA) workflows is often time-consuming. This study examines the effectiveness of different LLMs for diagnosing GA workflow failures. First, we introduce a novel approach for log processing that systematically filters and structures raw log data, significantly enhancing the quality of input for LLM. Second, we conduct a comprehensive evaluation of several enhanced prompt engineering strategies, including zero-shot, few-shot and self-refinement, to identify the most effective methods for generating accurate and actionable failure explanations. Our experiments with 100 GitHub GitHub Actions workflow logs demonstrate the promise of our proposed solution, showing substantial improvements in correctness, relevance, and clarity of failure diagnoses.