A question is not an authorization
People ask AI about a job change, a move, a purchase, or a relationship because a clean answer feels like relief. The model sees the text you provide, not every contract, obligation, health constraint, or family consequence behind it.
Treat the answer as a draft for thinking. The person asking still owns the decision, the evidence check, and the cost of being wrong.
Write three red lines before the prompt
Keep irreversible payments, signatures, and regulated professional decisions under human control. Do not submit another person's identifiable information without a lawful basis and careful redaction.
A third line is simple: when the evidence is incomplete, the next move must be stoppable. These boundaries prevent a fluent answer from quietly becoming an order.
Make the model argue back
Ask for four columns: known facts, assumptions, strongest case for, and strongest case against. Then ask which missing fact could change the recommendation and what would count as a stop signal.
This does not make the model infallible. It changes the interaction from seeking reassurance to locating uncertainty that a human can verify.
Run one bounded experiment
Interview three prospective users, compare two documented offers, or complete a small pilot that can be cancelled. Write the maximum loss and acceptance rule before sending the request.
On APIToken, use a separate project key, redacted samples, public status, usage records, and a budget ceiling. The model can help structure the test; it cannot approve your life.