Do I need to know how to code?

No. You describe what you want in plain English and Claude writes the code. You review it and run it. The two prompts above do all the setup for you.

Do I have to use the terminal?

Only once, for the two setup prompts, and you just paste them. After that you can build entirely in the desktop app with buttons.

The tool comes with a paid Claude plan. The free plan does not include it. Pro starts at about $20 a month. The 1 million word memory is automatic on a paid plan.

It is Claude Code's most powerful setting. It uses the deepest thinking and can run its own mini-team of helpers. You turn it on by typing /effort and picking ultracode.

For setting up your own files and building your own app, yes. If you want a safety net, use Auto mode instead. It runs hands-free but checks each step. Keep bypass for your own fresh project folders.

Do vibe coding apps really run on Claude?

Many do. Replit and Lovable are listed as Anthropic partners, and Emergent is an Anthropic customer built on Claude. You pay a markup to reach Claude through them.

This still feels too advanced. What do I do?

Install the desktop app, then paste the two prompts. They set up my whole system for you, step by step, and reassure you the whole way.

How to Vibe Code on Claude Code and Ship Real Apps

I burned more than two thousand dollars on vibe coding apps before I quit. Replit, Emergent, Lovable. My card got charged every few days and I barely noticed.

Then I learned the trick they hide, fixed my whole setup, and never looked back. Now I use this exact setup to build real systems and automations, and to ship my own apps. Two of them are live on the App Store and Play Store right now.

This is that setup. The real one I use every day, not a clean version for show. You will not find it written down anywhere else. And if any step below looks too advanced, relax. I hand you two prompts that set the whole thing up for you.

Why those apps drain your card

They feel like magic on the first prompt. Then you add one real feature and it breaks. You fix one thing and ten more break. Every fix spends more credits. That is the business model.

Here is the twist. Most of these apps run on Claude anyway. Replit and Lovable are listed as Anthropic partners. Emergent is an Anthropic customer built on Claude. You pay a markup to reach Claude through a middleman.

Anthropic's own partner list (Replit and Lovable are on it)

Their price grows with every prompt. Replit charges per checkpoint, and once your monthly credits run out it can bill your card with no cap.

Replit's own pricing page (charged per checkpoint)

Real users report charges of $50 to $100 every few days, and one builder posted over $1,000 in a single week. You are paying a middleman for Claude. So let us go straight to Claude.

Step 1: Install Claude Code (the desktop app)

You are not a coder, so use the desktop app, not the terminal. It has Claude Code built in, with buttons instead of commands. Download it and install it like any normal app.

Download the Claude desktop app

On Windows, install one small free tool first, called Git, then restart the app. Most Macs already have it. This is the only fiddly bit, and it is a two-minute click-through.

Git for Windows (install this first, then restart the app)

Open the app, sign in, and click the Code tab. You need a paid Claude plan. Start on Pro, about $20 a month. Move up to Max, about $100 a month, once you build every day. That $100 is what I pay, for the same Claude those apps charge you hundreds to reach. The huge 1 million word memory turns on by itself on a paid plan.

Claude plans and current prices

You should now see a Code tab with a Select folder button and a model dropdown. That is Claude Code. You did the hardest part, and you never touched a command. One trap to skip: do not set an ANTHROPIC_API_KEY. That turns on pay-per-use. Your plan already covers everything.

Step 2: My rulebook, the CLAUDE.md (this is the whole secret)

This next part is the real reason my apps stop breaking. Most people never do it. You are about to.

CLAUDE.md is a short rulebook Claude reads before every chat. It tells Claude how to think, how to work, and when to stop and ask you. I did not copy a generic one. I built mine by combining four things, then adding my own layer on top.

The community rulebook that grew from Andrej Karpathy's notes on where AI coding goes wrong. Over 187,000 people have saved it.
Karpathy's own LLM Wiki, his idea for how an AI should keep notes and memory as you work.
Anthropic's own official rules for a good CLAUDE.md, plus their Skills feature.
And my own Seven Laws that I run on every single system and app I build.

The community Karpathy rulebook (built from his notes, not by him)Karpathy's own LLM Wiki (his memory idea)Anthropic's official CLAUDE.md rules Anthropic's Skills (reusable powers for Claude)

Here are my Seven Laws in plain words. This is the heart of my whole setup. Read them once and you will get why my apps stop breaking.

My Seven Laws (the core of my rulebook)

# My rules for building with AI
1. Plan first. Show me the plan and wait for my ok before you write any code.
2. Make the smallest change that works. Do not touch anything I did not ask about.
3. Fix the real cause, not the symptom. Then tell me what changed and why it is safe.
4. One task, one fresh copy (a git branch). Never pile new work on finished work.
5. If you are unsure, stop and ask me. Never guess. Never make things up.
6. Read my real files before you change them. Check facts, do not go from memory.
7. On a big choice I cannot undo, show me the options first, then your pick and why.

That is the core. My full rulebook adds the exact stack I use, my design bar so nothing looks like cheap AI output, and how apps go live safely so I never break a real one. Good news: the first prompt below sets up the whole thing for you, so you do not have to write a word of it.

Step 3: Set up my whole system (one prompt)

Now the magic. This one prompt makes your Claude work exactly like mine. You run it once. It builds your rulebook and your memory for you.

First, open a terminal so Claude has full power for this one setup. In the desktop app you can use the built-in terminal, or on Windows open your Start menu, type terminal, and press Enter. Then start Claude with full power turned on by pasting this line:

start claude with full power (paste in the terminal)

claude --dangerously-skip-permissions --model best

That is bypass mode. It lets Claude do the setup without stopping to ask you each time. It is safe for this, since you are only setting up your own files. If you want a safety net that double-checks each step, swap in --permission-mode auto instead. It is Anthropic's newer, safer version and it still runs hands-free.

How the permission modes work (bypass vs auto)

Next, turn the brain up to full. Type /effort and pick ultracode. That is Claude's smartest, most powerful setting. Now paste the big prompt below and press Enter. Claude does the rest.

Prompt 1: set up my whole system (paste this)

You are my setup assistant. I am not a coder. Set up my whole build system on this computer, in plain words, one step at a time. Tell me exactly what I should see after each step. If a file already exists, do not overwrite it. Tell me and skip. Never delete anything or spend money without asking me. First ask me if I am on Windows or Mac so you use the right file paths, then go.

Do these three things in order.

1) MY RULEBOOK. Create the file ~/.claude/CLAUDE.md. This is the rulebook you read before every chat, on every project. Put the exact rules below in it, and tell me in one plain line what each part does. Keep it short. Anthropic's own guide on this file is at https://code.claude.com/docs/en/memory if you need it.

# My AI Operating System (read this before doing anything)
## Who you are, who I am
You are the builder. I am the boss, not a coder. I give goals, look at your plan, and say yes or no. Do not hand me raw command-line chores if there is a cleaner way.
## The stack we use every time
Next.js for the app, Supabase for the database and login, Vercel to put it online, GitHub to store the code. Default to a website. Only build a phone app if I clearly ask for one. If a project has a docs/PRD.md file, read it before you plan.
## The 7 Laws (never break these)
1. Plan before you build. For any 3-step or unclear task, show me a short plan first and wait for my go.
2. Smallest change that works. Do not rewrite a whole file unless I ask. Do not touch code that is not part of the task.
3. Fix the real cause, not the symptom. After a change tell me what changed, why it is safe, and what you left alone.
4. One task, one fresh branch. Never keep adding to a branch whose pull request already merged.
5. Stop when unsure. Do not guess, invent, or assume. My go means the exact plan I approved, nothing extra.
6. Research before you act. Read the real files first. Check facts from real sources, never from memory. Never say you checked something you did not.
7. Give me options on big, hard-to-undo choices. Show 2 or 3 real options with trade-offs, your pick, and why the others lose, before building. For small changes, just build.
## Design bar (any screen) so nothing looks like cheap AI output
Premium and mobile-first. Copy a named brand like Linear or Stripe, not "make it nice". No Inter, Roboto, or Arial headline font. No default purple-to-blue gradient. No fake names, numbers, or logos. No placeholder text or dead links. Build from shadcn/ui, then look at a screenshot and improve it two or three times before calling it done.
## How things go live (I control production)
You deploy free previews to Vercel and give me the link. You never push to the main copy, never merge pull requests, never go live to production. I click merge in GitHub when I am happy, and that ships it live.
## Never-degrade hygiene
One source of truth for every fact. No duplicate files, no dead files. Delete old ones when you replace them. Put experiments in a throwaway _scratch folder. Verify every price, date, and command. Never put passwords in git. End every change clean.
## Talk to me
Short, clear, honest about risk. Own mistakes without groveling. No em dashes. No AI-slop phrasing.
## When you get something wrong
If I correct you, ask if you should add a line to this file so it never repeats. This file grows only when a real lesson was learned.

2) MY MEMORY. Set up a simple second brain in plain markdown, using Andrej Karpathy's LLM Wiki idea from https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f, where I keep the real facts and you keep the notes. Make a folder called Brain with these files: index.md (a list that points to everything), log.md (an empty history, and you add one dated line when something important happens), a wiki folder (one short page per project), and schema.md with these rules: one topic per page, keep pages short and current, never store passwords, no em dashes. From now on you update these notes as we work, so a fresh chat always has the full story.

3) MY SAFETY. My rulebook above already says you never push to the main copy, never merge my own pull requests, and never go live to production, so I always click publish myself. Confirm you will follow that on every project.

When you are done, show me my CLAUDE.md and my Brain folder, and tell me in one sentence that from now on you follow these rules everywhere.

Rules for you the whole time: work in plain words, explain each step simply, reassure me when something looks scary but is normal, make the smallest change, ask before deleting anything or spending money, and stop and ask me the moment you are unsure.

You should now see a file called CLAUDE.md and a Brain folder. That is my whole system, living on your computer. From now on Claude follows these rules on every project you build.

Step 4: Install my toolbelt (one more prompt)

This second prompt installs the tools, skills, and connections I use, so Claude builds like a pro instead of a beginner. Paste it in the same window and let it work. It walks you through any logins.

Prompt 2: install my whole toolbelt (paste this)

You are my setup assistant. I am not a coder. Install my whole toolbelt so you are great at building real apps. Work in plain words, explain each thing in one line, and tell me what I should see. First ask me if I am on Windows or Mac so you use the right commands, then go. If something is already installed, skip it. Ask me before anything that needs a login or a paid key. If a command fails, tell me in plain words and try the backup I give you.

Install these, in order.

THE TOOLS (the building blocks):
- Node, the engine most apps run on. Windows: winget install OpenJS.NodeJS.LTS. Mac: brew install node.
- pnpm, a fast installer: corepack enable pnpm.
- The Vercel tool, to put apps online: npm i -g vercel.
- The GitHub tool, to save code. Windows: winget install --id GitHub.cli. Mac: brew install gh.
- The Supabase tool, for the database and login. Windows: scoop install supabase. Mac: brew install supabase/tap/supabase. If neither works, we just use npx supabase when needed.

YOUR SKILLS (extra powers for you):
- Vercel's official skills, best practices for fast, good-looking web apps: npx skills add vercel-labs/agent-skills.
- Supabase's official skills, for safe databases and logins: npx skills add supabase/agent-skills.
- Anthropic's official skills: run /plugin marketplace add anthropics/skills, then /plugin install document-skills@anthropic-agent-skills.

YOUR CONNECTIONS (so you can control real tools) add these:
- Supabase: claude mcp add --transport http supabase https://mcp.supabase.com/mcp
- Vercel: claude mcp add --transport http vercel https://mcp.vercel.com
- Sentry, which catches errors: claude mcp add --transport http sentry https://mcp.sentry.dev/mcp
- PostHog, which shows how people use the app: npx @posthog/wizard mcp add
- A browser, so you can see the app like I do: claude mcp add playwright npx @playwright/mcp@latest
- shadcn, for beautiful ready-made parts: later, inside a real project, run npx shadcn@latest mcp init --client claude.
Some of these open a login page the first time. Walk me through each login in plain words. Tell me what I should see when each one connects.

Rules for you: explain each item in one plain line, skip anything already installed, ask before any login or payment, use the backup command if the first one fails, and stop and ask me the moment anything looks unsure.

You should now have a full workshop: the tools to build, the skills to build well, and the connections to control real services. This is the same toolbelt I use every day.

Step 5: How you build anything from now on

Here is what your days look like now. You talk. It builds.

First, open ChatGPT or Claude and talk about your idea like you are talking to a friend. Keep going until it writes a clear plan for your app and saves it as one file. Call it PRD.md. That file is your blueprint. The more you perfect it, the better your app comes out.

Then make a new empty folder anywhere on your computer and drop your PRD.md inside it. Now start building, one of two ways.

On the desktop app: open that folder in the Code tab, set the mode to Bypass or Auto, pick the best model, set the effort to ultracode, then type @PRD.md and say build this whole app. (Turn on Bypass once in Settings, under Claude Code, if you do not see it.)

In a terminal: open a terminal inside the folder and run this, then type /effort and pick ultracode, then say: read PRD.md and build the whole app.

start a build (paste in the terminal)

claude --dangerously-skip-permissions --model best

From here it is all plain English. It writes the code, runs it, catches its own mistakes, fixes them, and shows you the real app. You look, you say what to change, and it changes it. When you love it, push it live on Vercel for free, or ship it to the app store.

Put your app online free on Vercel

You are not learning to code. You are directing a builder that never sleeps. The only limit left is how clearly you can describe the thing you want. That is how my apps reached the App Store and Play Store. Same Claude those apps rent to you, minus the markup.

If it breaks

You do not see Bypass or Auto mode in the desktop app: update the app (Help, then Check for Updates), then turn it on in Settings under Claude Code.
Claude keeps repeating a mistake: your CLAUDE.md got too long and the rule is buried. Make it shorter and add the word IMPORTANT.
Claude does too much at once: switch to Plan mode first so it shows a plan before it builds.
Your bill looks high: make sure ANTHROPIC_API_KEY is not set. Your plan already covers Claude Code.
The app will not run: ask Claude to run it and fix the error, and tell it to fix the real cause.
You feel stuck in a loop: type /clear, or press Escape twice to rewind, then start fresh with a clearer ask.

Do I need to know how to code?: No. You describe what you want in plain English and Claude writes the code. You review it and run it. The two prompts above do all the setup for you.
Do I have to use the terminal?: Only once, for the two setup prompts, and you just paste them. After that you can build entirely in the desktop app with buttons.
Is Claude Code free?: The tool comes with a paid Claude plan. The free plan does not include it. Pro starts at about $20 a month. The 1 million word memory is automatic on a paid plan.
What is ultracode?: It is Claude Code's most powerful setting. It uses the deepest thinking and can run its own mini-team of helpers. You turn it on by typing /effort and picking ultracode.
Is bypass mode safe?: For setting up your own files and building your own app, yes. If you want a safety net, use Auto mode instead. It runs hands-free but checks each step. Keep bypass for your own fresh project folders.
Do vibe coding apps really run on Claude?: Many do. Replit and Lovable are listed as Anthropic partners, and Emergent is an Anthropic customer built on Claude. You pay a markup to reach Claude through them.
This still feels too advanced. What do I do?: Install the desktop app, then paste the two prompts. They set up my whole system for you, step by step, and reassure you the whole way.

This free guide is the what. If you want the how, done with you, step by step, so you ship in your first lesson, that is my Claude for Beginners course. Want everything, agents, apps, and automations in one place? That is All-Access. And if you want me to set this up with you live and build your first real thing together, I take a few 1:1 people each month.