CO-STEP: A Prompt Engineering Framework Improving LLM's Response

Prompt engineering has become increasingly vital in enhancing the performance of Large Language Models (LLMs), as the quality of prompts significantly affects the accuracy and relevance of generated outputs. In this study, we propose “CO-STEP,” a novel prompt engineering framework designed to address the limitations of the existing “CO-STAR” framework, particularly in handling complex tasks requiring advanced reasoning. “CO-STEP” integrates clear goal setting, detailed context with examples, task breakdown, advanced techniques like chain-of-thought prompting, and iterative refinement to improve precision and adaptability. Evaluations conducted using the OpenAI 4o-mini model on benchmark datasets, including structured tasks like invoice data extraction and mathematical derivative computation, demonstrate the superior performance of “CO-STEP.” Specifically, “CO-STEP” achieves an accuracy of 0.98 and an F1-score of 0.97, reflecting its capability to produce coherent and task-aligned outputs. These results highlight the effectiveness of “CO-STEP” in optimizing LLM interactions, with ongoing empirical validation to further explore its applicability across diverse tasks.