Prohibitions are more enforceable than aspirations
“Be helpful and professional” cannot be checked. “Never quote a price” can be, automatically, on every output. The negative list is the part of the specification that a test suite can actually assert.
1. Commitments you can’t honour
Delivery dates, guarantees, availability, and outcomes. A promise made by your system is a promise made by you, and it is the class of failure that survives to become a complaint.
2. Prices and discounts
Unless pricing is fixed, published and current, quoting it invents a number you may have to honour. Where pricing is genuinely published, point at the page that states it rather than restating it from memory.
Anything it says that you would have to honour is something it should not be able to say.
3. Anything about a competitor
Comparisons generated on the fly can be inaccurate and occasionally defamatory. If you want a comparison, publish an honest one you control and have the system reference it.
4. Legal, medical, or financial advice
Regulated territory where a plausible answer creates real liability. These belong on the escalation list regardless of how confidently the model could respond.
5. Other customers, ever
Names, cases, or details about anyone else — the failure mode permissions at retrieval are meant to prevent, enforced again at the output layer as a second line.
6–7. Internal notes and claims about itself
Internal commentary must never surface, and it should not deny being an AI or invent a human backstory — disclosure holds under pressure or it is not a policy.
Write the list before launch, encode it as tests, and add to it after every incident. It is the most durable artefact in the whole deployment.