Design the AI wrong-answer path before the happy path
For any model-backed feature, the correction, dismissal, and override flows are specified as first-class requirements rather than edge cases.
How to run it
- 01
Write the mental-model statement: what the feature can do, what it cannot, and how confident it is.
- 02
Walk the Microsoft HAX playbook to enumerate concrete failure scenarios for your feature type.
- 03
Specify the correction path - how a user edits, rejects, or reverts model output - as an acceptance criterion.
- 04
Decide the confidence-display rule and hold to it; never present low-confidence output with high-confidence styling.
- 05
Test with users on deliberately wrong output, not only on the demo-quality result.
Anti-pattern this prevents
A feature demoed only on cherry-picked prompts, shipped without a correction path, and abandoned by users after the first confident wrong answer.