Student Gemini prompts, watched. Counselors told.
Vaulter pulls Gemini activity from Google Vault, scans every prompt against keyword rules, asks Gemini whether a match is a real concern, emails the flagged ones, and gives reviewers a queue to work. It is an Apps Script on a Google Sheet. Nothing to install, nothing to buy.
BOX 01HOW IT WORKSONLY RULE MATCHES THAT SURVIVE AI TRIAGE REACH A PERSON
Google Vaultholds Gemini activity for licensed students
Scheduled exportdaily, collected in the background
Keyword rulesprompts read in memory; no match, dropped
Gemini triagereal concern or not? likely false positives go to a log
Counselor reviewemail alert plus the dashboard queue
Gemini for Education is on for students and Vault keeps every prompt, but nobody reads Vault. Vaulter closes that gap: the first-person prompt about self-harm or the sextortion threat reaches a counselor the next morning. The same scan catches the chemistry student asking about a bomb calorimeter, and the AI step clears it before anyone's time is spent.
BOX 02LIVE DEMOTHE REAL DASHBOARD ON FICTIONAL SAMPLE DATA
Click a flag, run the bulk review, press Run import now. Every student here is invented.
Open the demo full screen
BOX 03WHAT A REVIEWER SEES
- A queue of flagged prompts, newest first, with severity Critical High Medium Low, category, grade and a two-line preview. Filters for unreviewed, severity and category, plus search.
- A detail pane with the full prompt and the matched keyword marked, the AI verdict and its reasoning, review notes, and a timeline: sent, flagged, alert emailed, AI verdict, reviewed by whom.
- Students are masked until the reviewer clicks Reveal, one flag at a time.
- Bulk review steps through everything unreviewed with the AI verdict beside each prompt and three choices: Confirm, Dismiss or Skip. Likely false positives clear themselves.
- Seven-day stats and a 14-day activity chart: flagged, unreviewed, critical open, share cleared by AI.
- Demo mode swaps in fictional sample data so the tool can be shown without exposing a real student. It is what you see above.
BOX 04PRIVACY
- Everything lives in the district's own Google Workspace: the script, the Sheet that holds flags, the alert email. No outside server, no vendor account.
- Only prompts that match a rule are kept. The rest are read in memory and discarded.
- Triage sends the matched prompt text to Google's Gemini API and keeps only the verdict.
- Vault is read as the admin who deploys the script. No service account, no key file.
- Dashboard access is a named list of district accounts with Admin and Viewer roles.
BOX 05WHAT YOU NEED
- Google Workspace for Education with Gemini on for students and Vault licenses for them.
- A super-admin with Vault privileges to deploy it. It runs as that account.
- A Gemini API key for triage. Optional: without it, matches are flagged without a verdict.
- A Google Sheet, the script from this repo, and an afternoon. The setup guide walks through it.
- Someone who will read the alerts. The tool moves a prompt to a person, nothing more.
BOX 06LIMITS, HONESTLY
- Keyword rules are blunt. "i want to die" also catches "i want to die my hair purple." The AI step clears most of these, but the rules need tending and the final call is a counselor's.
- It is not real time. Vault exports take minutes to hours, so the pipeline runs on a schedule and collects finished exports on the next pass.
- The AI verdict is triage, not a decision. Genuine Concern and Needs Review both wait for a person.
- It sees Gemini prompts only, not other chatbots, web search or documents.