The guardrails · Correction
When AI Gets It Wrong
Correct the source, not just the output
AI will get things wrong – a requirement misheard in a session, a component invented, a persona that doesn't ring true. The rule is simple. Correct the source, not just the output, so it doesn't happen again.
The Correction Loop
- 1SpotThe owner vets it, or a test fails
- 2FlagMarked on the artifact, visible to all
- 3CorrectThe owner fixes the output
- 4TeachFix the source it learned from
- 5ConfirmRegenerate, and check it holds
Where Each Lesson Goes
| When AI gets this wrong | Who corrects it | Where the lesson goes |
|---|---|---|
| Intent | CX and Product | The context files, so the next one is generated right |
| The Collab Brief | CX | The context files, and a note on how the session should be read |
| Code | Engineering, with the Definer | A new test, plus the guardrail or context file it broke |
| Designs and UI | UX | The design system AI project |
| Persona walkthroughs | UX, approved by CX | The personas themselves |
Never Silent
Every correction is logged, so the team can see where AI is weakest.
Never Twice
If the same mistake comes back, the lesson didn't land in the right source.
Reviewed Thursdays
Corrections from the week are part of hour 2 of the Collab, Thursday afternoon.