A Prompt Change Shouldn’t Move the Approval Boundary

A useful lesson from reasoning about AI workflow changes: a prompt revision can affect more than the wording of an answer—it can blur who is allowed to take the next step.

Consider an assistant that drafts a customer action but requires human approval before it proceeds. A focused regression case should check the whole transition: the draft is routed to the named reviewer, the action remains blocked while approval is pending, and an explicit approval is required to continue. Include a case where key information is missing, so the workflow has a clear pause or escalation route too.

That test can show whether these scenarios still behave as intended; it cannot prove every real-world case is safe. Keeping the prompt version and expected approval behavior together makes changes easier to inspect. The important question is not only “Did the response improve?” but “Did the right person still own the decision?”