Use this before changing a production prompt, system instruction, judge prompt, tool instruction, or retrieval instruction. A prompt change is a product change.
Upstream: behavioral contract from the AI PRD, quality bar from the eval plan, production signals from the observability plan. Downstream: approved changes feed into the weekly post-launch review.
Change summary
- Feature:
- Prompt or instruction changed:
- Current version:
- Proposed version:
- Owner:
- Date:
Why this change
- Trigger:
- Target behavior:
- Non-goals:
Behavioral contract check
- AI job still matches the PRD.
- Output contract is unchanged or PRD is updated.
- Safety boundaries are unchanged or risk register is updated.
- Human review and fallback behavior are unchanged or templates are updated.
Eval evidence
| Eval dimension | Baseline | Proposed | Pass? |
|---|---|---|---|
- Eval suite version:
- Golden set size:
- Regressions found:
- Manual review sample:
Rollout plan
- Staging only
- Shadow mode
- Canary
- Full rollout
Rollout notes
Rollback plan
- Rollback version:
- Rollback trigger:
- Who can roll back:
- Expected rollback time:
Monitoring
| Signal | Baseline | Alert threshold | Owner |
|---|---|---|---|
Decision
- Ship
- Hold
- Reject
Decision rationale