An experiment tested whether bad examples in session history can override skill-file rules for AI agents. Repeated incorrect patterns in prior exchanges caused the model to break its stated rules, even when corrected, suggesting session history can outweigh explicit instructions.