This example uses a temporary rule: “During the holiday period, in-stock orders are expected to ship within three business days.” Replace its dates, scope, and commitment with your real policy.

Prepare the rule#

Confirm who the rule applies to, its start and end dates, exceptions, and whether it requires live order or inventory data. Stable policy belongs in knowledge. A rule with triggers, time limits, or handling steps is usually better suited to an AI skill.

Create a learning proposal#

You can tell Yuna the durable rule directly. Yuna first asks whether it should continue organizing the instruction and creates a reviewable proposal only after confirmation. If the rule comes from an incorrect answer, select Correct AI in the Inbox and explain what was wrong, the correct handling, and its scope. Correct AI carries the customer question and original answer, but both entry points use the same suggestion, review, and testing flow.

Review the learning suggestion#

Check the referenced question, proposed answer, scope, expiry, and required capabilities. Narrow the scope and preserve human confirmation for higher-risk content such as refunds, prices, external commitments, or automatic sending.

Merge duplicate or equivalent suggestions into existing knowledge or skills to avoid conflicting versions.

Test before activation#

Test at least the original question, a rephrased version with the same intent, and an unrelated question that should not trigger the rule. Confirm that the answer does not broaden the commitment and that the rule is not used after expiry, without evidence, or when its conditions do not apply.

Activate only after review and Playground testing pass. Then spot-check the final answer received through a real channel.

Pause or roll back#

Pause the skill when the policy changes or an error is found so it is no longer used automatically. If a new version performs worse, roll back and rerun the key tests. Pause and rollback preserve history; deletion is not the first choice for correcting a rule.