Amey-Thakur Negative instructions
Do not mention pricing raises the salience of pricing. Negative instructions work less reliably than positive ones because they name the thing they forbid, and the fix is usually to state the desired behaviour instead.
Method
- Convert prohibitions to positives. Instead of do not use jargon, write use plain language a newcomer would understand, which is both clearer and more followable.
- Say what to do in the forbidden situation. If the answer is not in the sources, say so beats do not make things up, because it supplies the alternative behaviour.
- Keep genuine prohibitions few and prominent. Some things must be stated negatively, particularly safety rules, and they survive better when there are three rather than fifteen.
- Place hard prohibitions in the system prompt. They apply across turns and should not be overridable by user input (see system-prompt-design).
- Avoid naming the exact thing you fear. Detailed descriptions of unwanted output can act as a template for it.
- Verify rather than trust. For consequential prohibitions, check the output rather than relying on adherence (see agent-generate-and-verify).
- Test the negative case explicitly. Include prompts designed to trigger the forbidden behaviour in your evaluation set.
Boundaries
Reframing helps and does not eliminate the effect; models still occasionally do what they were told not to. Safety-critical prohibitions need enforcement outside the model. Some constraints have no positive formulation and must be stated as prohibitions.