Generative text-to-image AI tools allow users to generate images based on keywords, known as prompts, enabling users to create high-quality images. However, current research mainly focuses on professional prompt settings for text-to-image generation, with limited exploration of natural language prompts used by non-expert users. Therefore, this study aims to understand how non-expert users describe images using natural language and the challenges they face in the process, and then construct a prompt framework to help them more accurately generate expected images. Twenty non-expert users without art backgrounds were recruited to participate in an exploratory experiment, in which they were instructed to use natural language prompts to reproduce two types of given images using Midjourney. Based on the textual content of the collected prompts, effective prompts were extracted and subjected to hypothesis testing and directed content analysis. The results showed that participants tend to overestimate the similarity between the images they generated and the original images, suggesting that individuals tend to overlook certain aspects subjectively when unguided, which underscores the necessity for a systematic prompt framework. A natural language prompt framework containing five main themes (composition and proportion, color and tone, subject description, thematic expression, and details) was proposed, providing sub-entries and example prompts for each theme. The effectiveness of the prompt framework has been preliminarily verified. This framework can help users more effectively utilize text-to-image AI tools and improve the user experience of generative AI.