


Pick one AI feature you own. Define what a correct response should contain and which tool actions require approval. Build a representative evaluation set, recheck quality and fallback behavior after a change, then compare the result with your cost and latency limits. Record what you would do when an alert fires. Includes 12 task Skills and one router in a downloadable ZIP.
Set the output and action rules. Specify required outputs, who may authorize a tool action, and how a human review or withdrawal appears in the product.
Test a prompt or model change. Review representative cases, business regressions, latency, fallback behavior, and the evidence required to release.
Link quality to operating cost. Review browser instrumentation, alerts, incident recovery, unit cost, and peak capacity as separate checks.