1. Read the workflow and source
Open the complete free skill and authored Markdown download, its inputs and review gates. Read Claude project documentation for current file, plan and privacy limits. Account access is not tested here.
2. Start with approved reference files
Create a separate conversation or project for this one job. Upload the public Markdown as reference if your account supports files. This is not a plugin installation. Do not mix clients or place credentials and private records into a reusable setup instruction.
3. Run the fictional acceptance test
Fictional week: 9 refunds; 7 inside policy and 2 exceptions, one with documented evidence and one approved in chat with none. Expected result: flag the undocumented exception.
4. Use a narrow request
Use the attached Refund Policy Exception Decision Log as reference. Audit only these fictional refunds. Return request, policy clause, approver, evidence and verdict. Do not issue refunds or message customers.
5. Review before real data
Check each source row, unchanged original value, missing fact and arithmetic. A successful fictional result is a readiness check, not proof of real-account compatibility. Only then use authorized minimum-necessary real exports, confirm retention/access and ask the responsible owner to review the result.
Where this guide stops
No sending, sharing, ledger/CRM changes, account sign-in or spending is authorized by this guide. A separate request and the required approval govern real actions. Keep an evidence copy of the input version and reviewed output.
Checked November 1, 2026. Demand lead follows the published workflow topic; exact guide query volume is unmeasured. AI Agent Skills is independent of the named assistant maker.