Write the system prompt for a customer-facing assistant

The assistant handles {{what}} for {{who}}, on {{channel}}. It can access {{tools and data}}. Brand voice: {{describe, or paste three real examples}}.

Write the system prompt with these parts, in this order: identity and scope in two sentences, what it must always do, what it must never do, how to handle a request outside scope, tone rules with a good and bad example of each, and the exact escalation format.

Be specific where it counts: what it says when it doesn't know, what it says when the customer is angry, what it does when the tool returns nothing or something contradictory, and how it refers to competitors, pricing, refunds and legal questions.

Never let it invent policy. If a rule isn't in the prompt or the retrieved documents, the answer is escalation, not inference.

Then give me ten test messages designed to break it — including two that are reasonable but out of scope — and the response each should produce.