Stale Article Auditor
Score every knowledge base article for staleness - old dates, retired products, low ratings, and contradictions - then route the flagged ones through a human review gate before anything changes.
A private internal tool that scores and flags stale KB articles with plain-English reasons, lets a reviewer approve review / retire / keep per article, then exports an action list and notifies owners - never auto-unpublishing anything.
Before you start
- A KB export with title, body, last-updated date, and helpfulness score (CSV or Google Sheet)
- A simple list of your current vs retired products / features / prices
- Free accounts: Vercel, Supabase, Resend (the plan walks you through them)
The problem this kills
Your knowledge base started clean. Then products got renamed, prices changed, features were retired, and a few hundred articles quietly drifted out of date. Now agents don't trust the KB, customers get answers that are wrong, and "self-service deflection" stops deflecting anything. Nobody has time to re-read every article, so the rot just spreads.
The instinct is to bulk-unpublish anything old - but that is dangerous. A wrongly retired article kills deflection on a topic customers actually need. What you really want is a way to find the suspicious articles, explain why each one is suspect, and let a human make the call.
What you'll build
A small private web app, just for your team, that loads your KB export and scores every article for staleness using signals you can actually defend:
- Age - how long since it was last updated.
- Retired references - mentions of products, features, or prices you've marked as retired.
- Low helpfulness - articles your own readers rate poorly.
- Contradictions - older articles that disagree with newer ones on the same topic.
Each flagged article shows up in a review queue with its staleness score and the reasons it was flagged. A reviewer reads it and approves one action - review, retire, or keep. Nothing changes status automatically. When the review pass is done, the tool exports a clean action list and emails each article's owner what they need to do.
What's inside the Implementation Plan
- It interviews you first. Before building anything, the plan has the AI agent ask about your KB - how your export is shaped, what "retired" actually looks like in your data, your real product and SKU naming, your volumes, and your edge cases. Then it reads a short tailored spec back to you for a thumbs-up. You get a tool fit to your business, not a generic template.
- A step-by-step build you paste into an AI coding agent, one prompt at a time.
- A defensible scoring model you can tune - weights for age, retired references, low ratings, and contradictions.
- A reviewer queue with the reason for every flag, plus the human approve step.
- CSV / Google Sheet import and a clean CSV export of the full audit with recommended actions.
- Owner notification emails so the right person knows what to fix.
The governance it includes (this is the point)
- Login so only your team can open the tool.
- Row-level security so each organization only ever sees its own articles and audits.
- A full audit trail - who flagged, who reviewed, who approved which action, and when.
- A hard human-in-the-loop gate - the tool only ever recommends. A person approves review / retire / keep before any status is written. It never auto-unpublishes.
- Duplicate guards - the article ID is the dedupe key, so the same article can't be scored or actioned twice in a pass.
Who it's for
Knowledge managers and subject-matter experts (SMEs) responsible for content accuracy - anyone who owns a knowledge base and needs to trust what's published without re-reading the whole thing by hand.
You've got this. Open the Implementation Plan, paste the first prompt, and let the interview tailor the build to your KB.