This is annoying. Similar with "context poisoning". LLMs can be highly productive... until they're "poisoned" with contradicting inputs / guidance.