| Abstract: |
Large language model (LLM)-enabled applications increasingly rely on prompts embedded directly in source code. These prompts influence output quality, reliability, safety, security, response structure, and API cost, yet developers have limited support for inspecting them during development. This paper presents Prompt Radar, a Visual Studio Code extension that helps developers analyze prompts-as-code for potential prompt smells. Prompt Radar uses LLMs to assess prompts across five categories: Security, Safety, Reliability, Efficiency, and Format, and presents the results directly in the IDE through explanations and visual feedback. We evaluate Prompt Radar using real prompts mined from public GitHub repositories. A multi-LLM evaluation with ten models found weak agreement on exact smell scores but 92% agreement when scores differing by at most one level were considered a match. The results also show that Security is the most contested category, where developer review is especially important. To make the tool accessible, we provide a website with the GitHub repository, a usage demo video, and a survey for collecting feedback on usability, reliability, usefulness, and future improvements. |