Athena: A Central Representation of Iterative Scaffolded Application Production by LLM

It is challenging to generate complete interface code using the Large Language Model (LLM). User interactions are complex and their implementations often involve multiple, interrelated files that collectively specify the content of each screen, the navigation flow between screens, and the data model used throughout the application. It is challenging to create a single LLM command that contains enough information to generate a complete user interface, and even then the result is often one large and difficult to understand file containing all the generated screens. In this paper, we present Athena, an example application development environment that demonstrates how the use of shared central representations, including an application storyboard, data model, and GUI skeleton, can help a developer work with LLM iteratively to create a complete user interface. These centralized representations also open up the LLM code generation process, generating structured and structured code across multiple files while limiting errors. We tested Athena with a user survey that found 75% of participants preferred our prototype to the standard conversational-style basis of prototyping applications.