System prompt for an agent that stays in its lane

Write a production system prompt for an AI agent. You are an engineer who has shipped agents to real users and knows where they fail.

What the agent does: <its job in one sentence>
Who talks to it: <customers / employees / another system>
Tools it can call, with what each does: <list, or "none">
Information it has: <documents, database access, what it does NOT know>
Things it must never do: <e.g. promise refunds, give medical advice, reveal internal data>
When it should hand over to a human: <triggers>
Output format needed by the next step: <JSON schema / plain text / markdown>
Languages: <e.g. Hebrew and English, reply in the user's language>

Write the system prompt with these sections, in this order:
1. Role and scope, 3 sentences. What it is, what it does, what it does not do.
2. Knowledge boundaries: what it may state as fact, and the exact phrase to use when it doesn't know.
3. Tool use: when to call each tool, what to do with the result, and what to do if a tool fails.
4. Hard rules: the "never" list, phrased as unambiguous instructions.
5. Handover: the exact conditions and the exact message to the user when handing over.
6. Tone and language.
7. Output format, with one example.
8. Ten test messages, including adversarial ones, with the expected behaviour for each.

Under 600 words for the prompt itself. No vague instructions like "be helpful".