AI-assisted software development for brachytherapy seed inventory management.

PURPOSE: Medical physics departments frequently rely on in-house software to address niche clinical needs, yet these tools are often built by physicists without formal software engineering training, resulting in legacy applications that are difficult to maintain, poorly documented, and vulnerable to obsolescence from operating system upgrades. This study investigates the use of AI-driven "vibe coding", i.e. software development through natural language prompting, to enable a clinical physicist with no web development experience to replace an obsolete MATLAB-based radioactive seed inventory system with a modern desktop application. MATERIALS AND METHODS: Development was performed using Claude Code (Anthropic) powered by the Claude Sonnet 4.6 large language model. A structured "Plan-Review-Execute-Verify-Feedback" (PREVF) methodology governed all development, with the physicist providing natural language prompts and the AI generating all source code. The resulting application manages five seed models across four clinical applications, with features including seed batch lifecycle tracking, calibration record management, audit logging, waste deposit tracking, etc. Evaluation metrics included development time and prompt volume, feature completeness against eight pre-defined requirements, System Usability Scale (SUS) scores comparing the legacy and new systems, and an independent maintainability test in which team members with minimal coding experience attempted modifications without developer assistance. RESULTS: The application was developed in approximately 13 hours, using approximately 115 natural language prompts. All eight functional requirements were fully implemented. The new application received a median SUS score of 95 compared to 32.5 for the legacy system. In the maintainability evaluation, three independent testers completed bug fixes and feature modifications in one to two prompts each, all in under 10 min, without guidance from the original developer. CONCLUSION: AI-driven vibe coding enables clinical medical physicists to rapidly develop and maintain operational software, resolving the maintenance bottleneck characteristic of legacy in-house tools. The PREVF methodology provides a structured framework for physicist oversight of AI-generated code. This approach has potential applicability to provide clinical support tools by medical physics departments.