Build 40 Python projects where the instruction file asks for one convention and the working code uses the other, then let two mainstream coding agents add a feature. Across 320 unattended runs and 720 forced choices, 90.1% went with the stale file and 4.2% with the code. One agent named the conflict out loud in 79 of 80 runs and still followed the file in 86.2% of its choices. Add one sentence to the user-level instruction file and code-following goes to 91.4%, with the test pass rate moving from 89.4% to 100%.