Free vibe coding course: ship your first real app

A year ago vibe coding meant a chat window and a web builder. You described the app, it built a version, you patched it by hand.

Today the agent works inside a desktop app on your own computer. The Claude desktop app, the Codex app, Antigravity and Cursor read your whole project folder, build, test, deploy, and come back with a report. You direct. They build.

What this changes for you:

• You can ship real apps with auth, payments and a database, not demos.
• One brief can run for an hour without you.
• The web builders (Lovable, Bolt, base44) are still the fastest first hour. They are no longer where a real product lives.

What you need: a laptop, a $20 to $200 a month agent plan, and the habit of writing a clear brief. That is the whole course.

Two lanes. Choose one for your current project, not forever.

The desktop agent lane (the Claude desktop app, the Codex app, Antigravity, Cursor):

• You will ship to real users.
• You want auth, payments, a database and a deploy you own.
• You are fine with a project folder on your own computer.

The web builder lane (Lovable, Bolt, base44):

• You want a working prototype in an hour to show someone.
• You are validating an idea before building it properly.
• You will move to a desktop agent once it works.

The rule: if a stranger will pay for it, use a desktop agent. If you are still finding out whether anyone wants it, start in a web builder, then move.

Most members run both. Prototype on the web, rebuild in your agent app.

1. 1Download the Claude desktop app from claude.ai/download and sign in.
2. 2Open the Code tab and pick a folder for your project. An empty folder is fine.
3. 3Paste the prompt below as your first message.

Create a CLAUDE.md for this project. Ask me five questions about what I am building, who it is for, the stack I want, and how I want you to work. Write the answers into CLAUDE.md as rules you will follow in every session.

That file is your agent's memory. Every session reads it first.

Two habits that save hours:

• Start big work with: plan this first and show me the plan before you write code.
• End every session with: summarize what changed and what is still open into CLAUDE.md.

Using Codex instead? Install the Codex desktop app, open your folder and send the same first message. The file it reads is AGENTS.md. Antigravity and Cursor work the same way: open the folder, write the brief.

A brief has five parts. Paste the skeleton and fill it in.

Outcome: one sentence on what exists when this is done.
Who uses it: the person, what they are trying to do, on what device.
Constraints: stack, what must not change, what to reuse, budget of time.
Done means: a numbered list I can check. Include "typecheck and build pass" and "works at 375px wide".
Screenshots: attach the reference and say what to copy from it.

Plan first. Show me the plan. Wait for my yes.

Three things that make an agent ship instead of wander:

• A done list it can tick.
• A reference it can look at.
• Permission to ask one question before starting.

If your app calls an AI model, never parse prose. Ask the model for a shape and check the reply against it.

• Define the shape with zod.
• Send it as a JSON schema response format through OpenRouter, or the provider's own structured output.
• Validate the reply with the same schema. Reject it and retry on a mismatch.

The free Jev App Starter Kit does exactly this. Open it in your agent app, ask it to explain how it validates the model's reply, then ask it to copy the pattern into your project.

One more step the pros take: judge the output before you show it to anyone. That is what Jev is for.

My eight AI Employees are open source under MIT: GTM Engineer, SEO, Web Dev, Social Media, Ad Manager, Sales, Customer Satisfaction and Chief of Staff. Each one is a folder of scheduled routines that runs on your own machine on the agent you already use.

1. 1Download the repo and extract one role to a plain folder.
2. 2Open that folder in your agent app (the Claude desktop app, Codex, Antigravity or Cursor) and paste: Install the GTM Engineer from this folder.

It researches your business from your website, builds its dashboard, schedules its own routines and stops once to show you its first drafts.

Start with the GTM Engineer if you have something to launch. Start with SEO if you have a site nobody visits.

The three things most tutorials quit before. The house stack:

• Supabase for auth and the database. Row Level Security on every table, no exceptions.
• Stripe for payments. Checkout for buying, the Billing Portal for managing, webhooks for the truth.
• Vercel for hosting and scheduled jobs.

The trap that catches everyone: secrets live in three places. Your .env.local for local development, Vercel for the site, Supabase secrets for edge functions. One value can disagree with itself in all three and nothing errors.

List every environment variable this project reads, with the file and line. Then tell me which store each one must be set in (local, Vercel, Supabase) and which ones have a hardcoded fallback that would silently run if the secret is missing.

Three free downloads, in the order to try them:

1. 1Nano Banana Editor v2. A complete image editor app with a one click install. Ship it under your own name today, then read how it is built.
2. 2What is Jev? A five minute lesson on the AI model that decides instead of writes, with a real call and how to add it with an API key or through OpenRouter.
3. 3Jev App Starter Kit. The full app: Next.js, four working tools, a clean interface and a live demo. Open the folder in your agent app, say install and run this, then add a fifth tool with the prompt in its README.

Then post what you shipped in the Friday wins thread on Skool. Every win gets a reply from me.