These cases ask which operation the assistant may perform and which check belongs outside the prompt. Include the benign controls when judging a safety gate.
Safety lessons
- Request risk routing: classify the action before choosing a response
- Refusal contracts: decline the unsafe effect and preserve useful help
- Sensitive output gates: check the rendered answer before release
- System prompts: treat instructions as guidance, not a secret vault
- High-impact advice: route uncertain decisions to qualified review