Automate Content Triage with Human Approval

Goal

Build a content-triage flow that never publishes or deletes data automatically.

Prepare a test form, a review queue, and a workspace containing only simulated data.

Use a new form response or message as the trigger and read only title, body, and submission time.
Add deterministic rules to reject empty content, duplicates, and clearly invalid input first.
Ask AI for fixed JSON containing category, summary, risk flags, and confidence.
Validate the JSON; send parse failures and low-confidence results to a manual queue.
Write every result to a review table and create downstream work only after a person selects Approve.
Test normal, ambiguous, prompt-injection, and oversized inputs.

Every input has a visible classification record, and every external write requires human approval.

JSON is unstable: reduce fields and add schema validation.
Tasks are duplicated: store the trigger event ID and deduplicate before writing.
Sensitive information reaches the model: redact it before the AI step.