The FINISH kit

Is your AI-built app ready to launch?

Two agent skills, a secret scanner and a 55-item checklist that audit your app, then try to break it.

Free. MIT licensed. Works with Claude Code and Codex.

Install the skills in 2 minutes

Three steps. The installer copies local files only. No network, no sudo.

  1. Clone the repo

    git clone https://github.com/First-Mate-Technologies/ai-app-production-checklist
    cd ai-app-production-checklist
  2. Install the skills for your agent

    Claude Code
    bash scripts/install-skill.sh claude

    Installs both skills into ~/.claude/skills/

    Codex
    bash scripts/install-skill.sh codex

    Installs both skills into ~/.agents/skills/

    Add --project to install into the current folder instead, or --skill break-my-app for just one skill.

  3. Restart your agent, open your app's folder, and ask

No agent? Run the scanner.

Needs Node 20 or newer. Run it from your app's folder.

npx github:First-Mate-Technologies/ai-app-production-checklist .

Supabase? Paste the audit.

Paste sql/supabase-rls-audit.sql into the Supabase SQL editor and run it. It only reads system catalogs. Every row it returns is something to look at.

What you get back

Each prompt ends with a file you can read, share and work through.

You ask“is my app ready to launch?”
You getBlockers first, in PRODUCTION_READINESS.md
You ask“try to break my app”
You getTest cases, and a DEFECTS.md if you let it run them
You ask“check my Supabase RLS”
You getAn access review you can verify in your own database

What's inside

Four parts. Use one or all of them.

  • Agent skill

    Production readiness

    Ask "is my app ready to launch?" Your agent detects your stack, runs the scanner, checks 55 items against your real code, and writes PRODUCTION_READINESS.md with file:line evidence. Blockers come first.

  • Agent skill

    Break my app

    Ask "try to break my app." It writes requirements from your code, hands them to an independent tester, and files every failure in DEFECTS.md before anything gets fixed.

  • Secret scanner

    One command, no setup

    A small Node script that finds service-role keys in front-end code, committed .env files and unverified Stripe webhooks. No dependencies, no network, and it never prints a full secret.

    npx github:First-Mate-Technologies/ai-app-production-checklist .
  • SQL and checklist

    Supabase RLS audit and 55 checks

    One read-only query for the Supabase SQL editor finds tables with Row Level Security off and policies that trust user-editable data. CHECKLIST.md covers data, auth, payments, errors, tests, performance, deploys and launch day.

How the audit works

Everything is read-only. The agent shows you what it found, with evidence. You decide what gets fixed.

How the production readiness audit worksFive steps. Your code goes through the scanner, which looks for leaked keys and open tables. The agent then checks 55 checklist items and records file and line evidence. It writes PRODUCTION_READINESS.md with blockers first. You confirm which fixes to make before anything changes.1Your coderepo, routes,schema, env files2Scannerleaked keys,open tables3Checklist check55 items, each withfile:line evidence4ReportPRODUCTION_READINESS.mdblockers first5You confirm fixesnothing changeswithout your yes

Illustrative output, not a real app.

How “break my app” works

The method we use on our own products, packaged as a skill.

How the break my app skill worksSix steps. Requirements are derived from your app and written to REQUIREMENTS.md. An independent tester with a fresh context, on a different model when available, writes TEST_CASES.md, mostly edge cases. Running the tests is optional and needs your yes. Every failure is filed in DEFECTS.md before any fix.1Your approutes, forms,schema2RequirementsREQUIREMENTS.mdderived fromyour code3Independent testerfresh context,other model if it can4Test casesTEST_CASES.mdmostly edge cases5Run testsoptional, withyour yes6DefectsDEFECTS.mdfiled beforeany fix
9%
of the QueueMate test cases
were happy path

Why independent

The agent that writes your code tests what it thinks it built. So the tester should be someone else.

On QueueMate, a free restaurant queue system we built in days, our agents tested their own work during development. Then a different model wrote 554 test cases from the requirements. Only 49 of them, about 9%, were happy path. They surfaced 38 defects the builders' tests had missed, and we filed all 38 before a single fix.

Read: how we test AI-generated code before it ships →

Illustrative output, not a real app.

What it won't do

We would rather you know the limits now than find them the hard way.

  • It is not a security audit

    The scanner and checklist are heuristics. A clean scan means those mistakes are not there. It does not mean the app is ready.

  • It does not change your code without asking

    Read-only by default. It asks before it edits a file or touches your database. Break my app also asks before installing a test runner or running tests.

  • It never prints a full secret

    Findings name the file, the line and the kind of key. Values are masked, with at most the first four characters shown.

  • It leaves your database alone

    You paste the SQL audit yourself. The agent only connects if you hand it an explicit read-only connection.

  • It does not read your git history

    A key you committed once and deleted is still exposed. Rotate it. The scanner will not see it.

  • It cannot test a running app on its own

    Rate limits, race conditions and webhook retries need a running app. The checklist lists them, and break my app turns them into test cases.

FINISH kit: FAQ

Is it free?

Yes. The kit is MIT licensed. Clone it, copy it, hand it to your team.

Does my code leave my machine?

The scanner runs locally and makes no network calls. The skills run inside the AI agent you already use, so your code goes only where it already goes when you use that agent. First Mate never receives it.

Which stacks does it work with?

It is built for apps made with Lovable, Bolt, Replit, Cursor and v0, usually React or Next.js with Supabase and Stripe. The scanner checks Supabase and Stripe patterns. The checklist and the skills apply to any backend, and items that do not apply are marked N/A.

What if it finds a lot?

Start with the blockers. The report lists them first, each with the file, the line, why it matters and a concrete fix. Work the first three checklist sections (data, auth, payments) before anything else. If it is more than you want to take on, we can help.

Do I need Claude Code or Codex?

No. The scanner is a Node script (Node 20 or newer), the SQL audit is a paste into the Supabase editor, and the checklist is a Markdown file. Only the two skills need an agent.

Want a second pair of eyes?

Thirty minutes on a call, free. Or send us the URL and an engineer emails you one finding within two business days.

Get started

Start your two-week trial.

No long-term commitment. Most engagements start with a low-risk trial. Prefer to talk first? Reach us directly.

By submitting, you agree we may contact you about our services. Unsubscribe anytime. See our Privacy Policy.