Say what to do instead of what not to do
prompting basics
// what it does
Negative rules ("do not ask for personal info") tell the model what to avoid but not what to do instead, so it often stalls or improvises. Replacing each "don't" with a positive next action gives it a path to follow. OpenAI's best-practices guide calls this out directly with a customer-support example.
// prompt
> You are a support agent. Diagnose the customer's login problem and suggest a fix. Do not ask for personal information such as username or password; instead, point them to the reset guide at [help URL]. Customer: [paste message]// gotcha
A prompt stacked with "do not" rules can backfire — the model fixates on the forbidden thing or freezes. Rewrite prohibitions as the action you want: instead of "don't ask for their password," say "refer them to the reset page at [link]."