Setting the guardrails
Approvals, explicit limits and human review to stay in control.
ObjectiveBy the end of this lesson, you will be able to place approvals, limits, allowed folders and human review around an assistant, then put them to the test with trick questions.
A guardrail is a rule that protects even when the instructions have been misunderstood. Learnya provides a first, built-in one: any action that commits you, such as sending, buying, publishing or deleting, waits for an approval. Your job is to decide where to place the others.
The guardrails at your disposal
- Approvals: decide how long they last, once, for the session or for the project.
- Limits in the instructions: excluded topics, people never to contact, tone to avoid.
- Allowed folders: for mail, the agent only files into the folders you allow.
- Console restrictions, set by the administrator for the whole organisation.
- Human review: a person approves the result before it leaves the team.
Sensitive tools never benefit from a broad approval. They ask again at the start of every session, even if you chose “Approve for this project”. This is by design: what cannot be undone always deserves a look.
Try it in LearnyaReview the instructions of this assistant and suggest guardrails: the situations where it must stop and ask, the topics it must decline and who it should refer the person to. Present them as sentences to add to the instructions. Try in Learnya
- Add a “Limits” section to the instructions with the sentences you kept.
- Test an out-of-scope question and check that the assistant refers to the right person.
- Ask for an action that commits you and check that the approval appears.
- Open the “Activity” screen to review the history of approvals.
- Decide with the team which approvals can apply to the whole project.