Exploring Prompt Engineering for Generative AI-Based App Generation

Abstract

We introduce a cutting-edge learning platform powered by large language models that enables students to effortlessly generate mobile applications for smartphones and tablets from natural language descriptions. We further demonstrate that these user-generated apps can be further optimized with minor adjustments to the generative model's input, or, its "prompt." To maximize the efficacy of the prompt in producing a desired application, we explore three different methods of modification: 1) altering the selection mechanism of example pairs, 2) varying the number of example pairs, and 3) changing the order of pairs within the prompt. The prompts are constructed from a collection of example pairs, which comprise a textual description of an example app and its corresponding code, in addition to a description of the desired app. We test the model's performance by evaluating it with 18 different mobile application task descriptions, ranging from basic to complex, and then leveraging BLEU score to compare the model's outputs to manually created apps. Our findings indicate that the method of determining example pair selection and varying the number of examples included can significantly influence the quality of the generated apps. However, reordering the placement of the example pairs within the prompt does not affect the outcome. Finally, we conclude with a discussion on the potential implications for computer science education. The platform we present in this paper aims to further the democratization of app creation through enabling users to create apps with ease, regardless of their technical background.