Unveiling Power on Combining Prompt Engineering Techniques: An Experimental Evaluation on Code Generation

Prompt engineering techniques have seen a significant rise in research interest as a means to achieve satisfactory results without retraining Language Models. This work presents a set of experiments to analyze the power of a combination of prompts. To this end, it evaluates six prompt techniques, combining them to result in twelve experimental scenarios applied to Python code generation. Evaluation using BERTScore indicates that Role combined with RAG achieves the highest performance in code generation with 98% similarity. Skeleton-of-Thought and Self-Verification reveal promising opportunities for the design of prompt templates. Our findings contribute to unveiling the power of combining prompt techniques for current applications such as code generation.