Part 7 of 14 · 1 min read
A refusal should help the model do the next right thing
A good refusal says what was wrong and what to do, without leaking how the system is built.
When the gateway says no, the model reads the refusal text and decides what to do next. A refusal that just says "denied" teaches it nothing; a refusal that explains the limit and suggests a next step gives it something to act on.
| Call | Refusal text |
|---|---|
get_ticket(6) for another customer |
Error: ticket 6 is not part of this conversation's customer. |
get_payment(2) without the permission |
Error: not permitted. get_payment needs the 'payments:read' permission, which this task was not given. |
refund(2, 5000) |
Error: refunds above 50.00 need a human agent. Offer the customer a hand-off. |
| a person says no | Error: the person did not approve this action. Do not retry it; tell them what you wanted to do and why. |
delete_everything() |
Error: unknown tool 'delete_everything'. Available: get_ticket, list_open_tickets, add_note, ... |
Compare this with a raw database error or a stack trace, which tells a model nothing useful and tells an attacker how the system is built. Module 6 covers the same idea for an MCP server: tool errors should tell the model what to do next, not just that something failed.