The Vibe Check

The Vibe Check Skill

Vibe coding is real and good: an idea becomes a working app in a weekend. Then the app gets users, and the things that didn't matter Friday night matter enormously — the API key sitting in client code, the endpoint that trusts the browser to say who's logged in, the database where every user can read every row. This skill is the bridge from "it works" to "strangers can use it": a structured self-audit ordered by embarrassment-per-fix, honest about what must block launch versus what can wait for week two. It reviews the user's own app — it's a seatbelt, not a lockpick.

What This Skill Produces

• A ranked findings list: 🔴 launch-blockers / 🟡 week-one / 🟢 eventually, each with the concrete fix (and the code-level change where code was shared)
• The attacker's tour: "here's what I'd try first on your app" — the 10- minute walkthrough of your own front door, as motivation and test plan
• A launch checklist for this specific stack, not a generic OWASP dump
• The data honesty check: what you're storing, whether you need it, and whether you can delete it when a user asks

Required Inputs

Ask for (if not already provided):

• What the app does and who's about to use it (5 friends? the internet? payments? minors? — the bar moves)
• The stack: framework, hosting, database, auth approach, AI APIs used
• Access to look: the repo/key files pasted, or answers to the checklist questions honestly ("is the Supabase anon key doing all your auth? be honest")
• What the AI assistant built vs what the user wrote/reviewed — unreviewed generated code is where the holes cluster

Framework: the five embarrassing holes (check these first)

1. Secrets in the client. API keys in frontend JS, .env committed to the repo, keys in the mobile bundle. Fix: server-side proxy for anything with a bill or a scope; rotate anything that ever shipped to a browser — it's burned, rotation is not optional.
2. Auth theater. The UI hides the admin button but the endpoint answers anyone; user ID taken from the request body; "logged in" checked in React but not on the server. Fix: every endpoint re-checks identity and authorization server-side; the client is a rumor, not a witness.
3. The database trusts everyone. Default-open row-level security, every user can query every row, the AI wrote select * where it meant where user_id =. Fix: RLS/scoped queries, then test as a second user — the two-account test finds most of it.
4. Input goes straight in. Unvalidated input into queries, prompts (injection into your LLM calls — your system prompt is not a secret once users can talk to it), file uploads with no limits, HTML rendered unescaped. Fix: validate server-side, parameterize, cap sizes, escape output, treat LLM output shown to other users as untrusted input too.
5. Data with no exit. Storing more than needed, no deletion path, AI conversation logs kept forever by default, no answer to "delete my account." Fix: store less, add the delete path now (retrofitting it after growth is 10x the work), write the three-line privacy note that matches reality.

Then the supporting cast: rate limits on anything that costs money per call (your AI endpoints especially — one loop = one invoice) · error messages that don't leak stack traces · dependency audit (npm audit is free) · backups tested once · the bus factor file (how to deploy, where the keys live).