Off-topic deflection
Detects messages outside the scope of customer support — political opinions, jailbreak attempts, unrelated requests — and declines politely while keeping the door open for in-scope questions.
Use as a guardrail wrapped around any customer-facing workflow. It prevents the worst public-facing AI failures.
It does not lecture, does not moralize, and does not engage substantively with the off-topic content. It declines, names the scope, and stops.
- intake
- classify
- decline
Provisioning the MCP URL for Off-topic deflection is part of Pro.
Free accounts can read the YAML, run the Live Preview, and provision one starter URL. Pro provisions URLs for the whole catalog.
Open YAML. Read it before you subscribe.
The harness that enforces this YAML is closed and hosted. The YAML is what you can audit, fork, and reason about.
1# off-topic-deflection.yaml2name: off-topic-deflection3version: 1.0.04category: compliance5description: Politely declines off-topic. No engagement.6 7scope:8 in_scope:9 - product_support10 - billing11 - account12 out_of_scope:13 - politics14 - opinions15 - jailbreaks16 - unrelated_advice17 18harness:19 - step: classify20 against: scope21 - step: decline22 say: scope_decline_message23 suggest: in_scope_examples24 - step: close25 26policy:27 lecture: false28 moralize: false29 engage_substantively: false30 Need off-topic deflectiontuned to one client’s voice, escalation logic, and required fields? That’s a $99 personalization.
One-shot fee. Turnaround stated up front. The result is yours, with the deterministic harness already wired.