You're tuning a customer-service LLM to stop suggesting refunds in cases where refunds aren't allowed. Which directive is more reliably effective?

Models follow 'do this' instructions more reliably than 'don't do that' instructions. The reason has to do with how the model processes the prompt. A positive directive maps directly to an action. The model reads 'when X happens, do Y' and treats Y as the thing to generate. A negative directive is harder. The model has to recognize the trigger, model the prohibited action, then try to avoid it. That extra step is fragile. Sometimes the suppression works, sometimes the prohibited action slips through. For the refund case, the negative 'do NOT suggest refunds' is the anti-pattern. Sometimes the model complies. Sometimes it suggests a refund anyway and then says it cannot. The positive 'for requests outside our refund policy, redirect the user to file a formal complaint via the support portal' gives the model a clear action with a clear trigger. The pattern: every prohibition can be reframed as a substitute behavior. Find the action the model should take instead, and state that.

Positive directives outperform negative ones because they give the model an action to execute rather than an action to suppress. This is one of the most replicated findings in instruction-tuning literature and a routine prompt-engineering best practice. The mechanism is straightforward. A positive instruction 'when X, do Y' establishes a clean attention path from trigger to action. A negative instruction 'do not Y' requires the model to attend to Y, model what Y would look like in this context, and then suppress it. That suppression step is fragile and inconsistent across inputs. For the refund example, the right framing is option B. The wording 'For requests outside our refund policy, redirect the user to file a formal complaint via our support portal' gives the model a recognition condition and a substitute action. Compliance jumps because the model no longer needs to suppress anything; it just executes the substitute. The two anti-patterns are repetition of the negation, which can prime the prohibited behavior, and stronger language like FORBIDDEN, which does not address the underlying attention problem. The fix is structural: replace prohibitions with substitute behaviors.