Title: Prompt Management Version: 1.0 Owner: [TENANT_CONFIGURATION_REQUIRED — AI Governance Lead] Status: Draft Last reviewed: 2026-09-07 Next review: [TENANT_CONFIGURATION_REQUIRED] Reviewers: AI Governance, Security
Defines lifecycle management for system prompts stored under prompts/system/, prompts/skills/, prompts/evaluation/, and prompts/red-team/.
Every prompt file has a version tag recorded in the prompt_version table (data-dictionary.md) and referenced by match_score.prompt_version_id and equivalent fields elsewhere, so every AI output is traceable to the exact prompt that produced it.
flowchart LR
A[Draft prompt change] --> B[Peer review - AI Governance + domain owner]
B --> C[Run against evaluation suite - ai-evaluation-strategy.md]
C --> D{Pass thresholds incl. fairness + red-team?}
D -- No --> A
D -- Yes --> E[Approve + version tag]
E --> F[Stage rollout - canary tenant/traffic %]
F --> G{Staged metrics healthy?}
G -- No --> H[Rollback to previous version]
G -- Yes --> I[Full rollout]
H --> A
Each prompt file has a designated owner (skill owner) and requires AI Governance Lead sign-off before promotion to production, recorded as approved_by/approved_at on the prompt_version record.
evals/datasets/ and evals/cases/ (see ai-evaluation-strategy.md).Rollback is a configuration change (pointing the skill’s active prompt_version_id back to the prior version), not a code deploy — enabling fast recovery from a regression. Rollback events are logged and trigger a mandatory post-incident note if the regression reached production traffic.
| Version | Date | Author | Change |
|---|---|---|---|
| 1.0 | 2026-09-07 | Documentation package generation | Initial creation |