AI Backlog Generation from App Mockups Evaluated
A new study on arXiv (2607.22902) presents an innovative way to develop sprint backlogs from visual app mockups, which are early-stage project tools. The researchers experimented with three different prompting strategies using GPT-4o: a zero-shot baseline, a Compositional Chain-of-Thought (CCoT) method for combining vision and language, and a persona-driven prompt. They analyzed seven app development projects from two countries and conducted interviews with developers. The baseline method focused more on recall than precision, while CCoT achieved average F1 scores between 52-66% for epics and user stories. While generating tasks was tricky, adding architectural context increased precision by up to 35%, especially for backend tasks, alongside valuable feedback from developer interviews.
Key facts
- arXiv:2607.22902v1
- multimodal approach for backlog generation from app mockups
- three prompting strategies on GPT-4o: zero-shot, CCoT, persona-driven
- seven app development projects across two countries
- CCoT achieved F1 scores of 52-66% for epics and user stories
- precision gains up to 35% with architectural context for backend tasks
- baseline prompt favors recall over precision
- tasks were more challenging to generate accurately
Entities
Institutions
- arXiv