Where the work starts
Separate return, shipping and product queries from fictional inbound messages.
Start with a fictional example that resembles normal work, but remove identifiable customer, employee or business information.
Prepare reliable order-status answers without assuming a shipment has been delivered.
Separate return, shipping and product queries from fictional inbound messages.
Start with a fictional example that resembles normal work, but remove identifiable customer, employee or business information.
Draft a response asking for the minimum necessary order details.
The output is a proposed working document, not a verified record or a completed business action.
ACT AS: a careful internal drafting assistant supporting ecommerce support advisers. GOAL: Separate return, shipping and product queries from fictional inbound messages. PRODUCE: Draft a response asking for the minimum necessary order details. INPUT: [replace with a fictional scenario containing a few deliberate missing details]. DO NOT: assume facts, make commitments, reveal personal information or act without approval. When uncertain, explain what must be checked. REVIEW RULE: Verify actual order status and refund eligibility in approved systems. FORMAT: (1) request summary, (2) essential missing details, (3) draft for a staff member to review, (4) safety or uncertainty notes.
Never copy confidential records into a service your employer has not approved. Treat AI drafts as untrusted until someone checks them.
Test a missing order number, split delivery and disputed return.
Repeat with one missing field and one deliberately contradictory detail. The system should highlight ambiguity rather than conceal it.
Verify actual order status and refund eligibility in approved systems.
Record who has the authority to review the result, correct errors and decide whether the draft can be used.
Track customer contacts resolved without unnecessary follow-up. using a small fictional or securely controlled sample. Count the human effort required to verify the result, the number of corrections and the cases where the draft fails. If the checks take longer than the original task, simplify the process instead of automating more.