1. Define the retrieval scope and consequence
Write down the first Rovo use case, the Confluence spaces it may draw from, and the cost of a wrong answer. A support refund policy, security runbook, and social event page should not receive the same risk treatment.
- Name the first user group and the decision they will make from the answer.
- List the spaces and permission boundaries involved.
- Mark high-consequence topics such as approvals, security, pricing, refunds, and incident response.
- Choose one space for the first evidence-based review.
2. Check content trust signals
Freshness is necessary but not sufficient. Two recently edited pages can still contradict each other. Review page age together with ownership, duplicated guidance, deprecation markers, evidence quality, and prompt-like instructions embedded in content.
- Find stale and outdated guidance using thresholds appropriate to the topic.
- Compare duplicated policies and runbooks for conflicting claims.
- Identify ownerless pages and assign a reviewer, not just a last editor.
- Label deprecated pages and make the current source of truth explicit.
- Review prompt-injection-like content as untrusted input, not as an instruction to the app.
3. Convert findings into governance
A one-off cleanup decays quickly. Each accepted finding should lead to an owner, a due date, and a clear action: update, merge, deprecate, restrict, or document as an accepted exception.
Measure the workflow using completion and recurrence, not only a single health score. Re-scan changed pages and high-consequence spaces on a cadence the team can sustain.
- Require human review before creating or closing cleanup tickets.
- Keep evidence short enough to minimize data while remaining understandable.
- Track open high-severity findings and time to resolution.
- Revisit scope, permissions, and source-of-truth pages when the Rovo use case expands.