Type: url
Original: https://arxiv.org/pdf/2606.12231
Status: generated
Summary
A mixed-methods empirical study of Rules, the configuration artifact AI IDEs use to inject persistent project-specific constraints into an LLM’s context. Mining 7,310 rules from 325 rule files across 83 open-source projects built with Cursor, Windsurf, Trae, Qoder and Kiro, the authors build a taxonomy of 5 primary and 25 secondary categories and triangulate it against a survey of 99 practitioners. They report a gap between perceived priority and actual configuration — developers rate architectural rules highest while repositories are dominated by workflow and formatting rules — and find that updating rules raises average artifact compliance by 22.99%, from 49.14% to 72.13%.