The authors built 40 Python projects with deliberately mismatched conventions and ran feature-addition tasks 320 times.
Without an added precedence rule, 90.1% of agent choices followed the stale instruction file. Claude Code identified the conflict in 79 of 80 runs but still followed the file. Codex CLI rewrote existing code to match the file in 20 of 80 runs; project tests failed in 17 of those rewrites.
The experiment added one sentence to a user-level instruction file: when repository documentation conflicts with existing code, follow the code conventions and report each discrepancy. The reported share of choices following code rose from 4.2% to 91.4%; rewrites stopped and the test-pass rate rose from 89.4% to 100%.
This is an arXiv preprint based on constructed Python projects. It does not establish that the same rates apply to production repositories, other languages, or other agent versions. Source: https://arxiv.org/abs/2610.30571