Design rules make project-specific design decisions explicit and checkable, helping teams preserve architectural intent as systems evolve. Although automated checkers can detect violations, developers must still interpret the rule, locate the code that must change, and construct a repair consistent with project conventions. We present DesignFix, an IDE-integrated approach that turns checker-reported violations into project-grounded, reviewable repairs. DesignFix augments each violation with the rule description, a rule-satisfying example from the same project, and focused structural context, while a two-step LLM pipeline separates repair reasoning from structured edit extraction. We report three complementary evaluations. A systematic study of seven prompt configurations across 9,450 repair trials found that the full workflow achieved up to 90% repair success, 89% explanation correctness, and 90% preservation of unrelated code: 26 percentage points above a baseline prompt. We conducted a controlled experiment with 14 developers. Compared to a control condition with browser-based LLMs, developers completed repairs 36% faster with DesignFix, increased correct repairs from 43% to 93%, and reported significantly lower workload on five of six NASA-TLX dimensions. Finally, we ran the autonomous in-IDE agent Cline on six repair task types drawn from our user study. Cline produced a correct repair for only half of them, while averaging 12.1× DesignFix’s token use and substantially higher API cost per run. Together, these results suggest that effective design-rule repair depends on grounding the LLM in checker-derived, project-specific design intent and on keeping repairs under developer review, needs that repository autonomy alone does not meet.

@inproceedings{paperbot12249,
  title = {DesignFix: Context-Augmented Repair of Design Rule Violations},
  author = {Mainul Hossain and Thomas D. LaToza},
  booktitle = {ASE},
  year = {2026},
  doi = {10.1145/3832783.3837422}
}