An AI interface must reveal what the system can access, what it proposes, and what a person is approving. Design the review and recovery states with the same care as the generated answer.
Show scope and provenance
Separate user instructions, retrieved material, and model output. Treat retrieved instructions as untrusted content. Identify the source and freshness of evidence, and explain important unknowns without inventing a confidence percentage.
Review consequential actions
Present affected resources and expected effects before a write. Bind approval to the exact reviewed plan; editing the plan invalidates approval. Authorization belongs on the server, not only in a disabled button.
Support interruption and recovery
Provide cancel, retry, partial-result, and failure states. Be honest about what already happened after cancellation. Keep an audit trail and offer undo only when a real reversal exists. Do not imply a demo is connected to an AI model.
Default decisions
Starting conventions for this system. Validate them in the context of your product.
- Default capability
- Least privilege
- Approval
- Bound to a reviewed plan version
- Evidence
- Sources, scope, freshness, limitations
- Failure
- Preserve user input and explain recovery