Vibe coding: what it is, tools and limits

Vibe coding, explained simply

So what is vibe coding? It's a way to build software by describing it to an AI instead of writing the code yourself. You say what you want in plain language, the AI writes the code, you look at the result and you refine your request. Someone who works this way is called a vibe coder.

For a mobile app, it can start with "create a to-do list with a button to add an item." But the important part comes next: checking that adding works, that the list is still there after you close the app, and that the screen is usable on your phone.

You can start without knowing how to program. To keep going, you'll gradually need to understand what your app does, where it stores data and how to check a change.

I build my own iOS apps with Claude and Cursor, every day. Solys, an alarm clock with mini-games, and TradaZen, a trading simulator, are on the App Store. This guide comes from that work: what vibe coding really lets you do, where it breaks, and how to get all the way to a published app.

Why is it called "vibe coding"?

The term has a birth date. On February 2, 2025, Andrej Karpathy, cofounder of OpenAI, described a new way of programming on X: you give in to the "vibes," accept what the AI suggests, and almost forget the code exists. He called it "vibe coding." The history of the word is laid out on its Wikipedia page.

The word caught on fast. Merriam-Webster listed it as early as March 2025, and the Collins dictionary made it its word of the year for 2025.

In France, curiosity followed. In Google Keyword Planner, the query "vibe coding" showed, on October 2, 2026, between 10,000 and 100,000 searches per month and a +900% change over one year.

Remember the original meaning: when Karpathy says vibe coding, he means no longer reading the code. That's great for a weekend experiment. It's risky for an app real people are going to use, and that's the whole difference with the method described below.

Vibe coding, no-code and AI-assisted development

These terms overlap, but they don't mean exactly the same method.

Method	What you do	What you get	The limit
No-code	You assemble blocks in a visual interface	An app that lives inside the tool	You depend on the platform and its pricing
Vibe coding	You describe the result and move forward with prompts	Real code that you barely read, if at all	Mistakes slip through as long as nobody checks
AI-assisted development	You also describe, but with rules, review and tests	Real code that you know how to evolve	Slower at first, you have to learn to verify

You can start with the second approach and gradually adopt the third. The tool's name isn't enough to tell them apart: you can give Codex a vague change request just as you can write a precise request in an app generator.

The guide to building an app without coding covers what you need to learn once your prototype starts turning into a real product.

How do you vibe code?

The method comes down to five moves, whatever the tool.

1. Describe a result you can check. Not "a fitness app," but "a screen with the list of my workouts and a button to add one."
2. Ask for one thing at a time. One screen, one feature, one fix. Ten requests in one prompt give you ten half-done things.
3. Look at the result on a real phone. A browser preview proves nothing for a mobile app.
4. Fix with a precise request. Say what you did, what you saw and what you expected. A screenshot helps a lot.
5. Save what works. Before each new request, keep a version you can go back to.

These five moves are enough for a prototype. For an app you want to publish, you need to add rules and checks: that's the topic of building an app with AI.

Which tools should you use for a mobile app?

There are two possible paths: work in your own project, or start in a tool that generates an app from your description.

Codex and Claude Code can work in a code folder, edit files and run commands. Cursor lets you work in an editor with an agent. These fit an app whose project you want to understand and keep. Their capabilities are described in the Codex CLI, Claude Code and Cursor Agent docs.

Generators have their own environments and output formats. Before picking one, check whether it produces a web app, a mobile project, or a project meant only for Apple platforms. The Lovable, Rork and FlutterFlow comparison explains that difference.

Here's what four well-known tools produce and what their free plan allows, based on their pricing pages read on October 2, 2026. These plans change often: check them before choosing.

Tool	What it produces	Free plan
Claude Code	Code in your own project, from the terminal	No: it's included in paid plans, starting at $20 per month
Cursor	Code in your own project, from an editor	Yes, with a limited number of agent requests
Lovable	A web app	Yes: 5 credits per day, 30 per month max
Bolt	A website or web app, hosted on its platform	Yes: 300,000 tokens per day, 1 million per month, with its logo

For an iPhone app, the project format decides what comes next. An agent working in a React Native and Expo project leaves you the code and the path to the App Store. A web app generator gives you a link to share, not an App Store listing. To choose, read which AI to use to build an app.

Above all, pick a workflow you can actually follow: open the project, display the result, save a version and understand an error. An impressive tool that leaves you stuck at the first fix won't help you finish.

What are the dangers of vibe coding?

The danger isn't the AI. It's publishing what nobody checked. Two public incidents show it, and so do my own Apple rejections.

Data open to everyone. A security researcher analyzed 1,645 projects built with Lovable and published his findings in May 2025: 170 were exposing data, including emails, phone numbers and API keys. The cause: database access rules that were missing or too broad. The app worked perfectly on screen, and that's the trap.

A deleted database. In July 2025, Replit's agent deleted a project's production database during a code freeze, even though its owner had forbidden any changes. Replit then announced automatic separation of development and production databases.

Apple rejections. The AI doesn't look at your app through a reviewer's eyes. On Solys, I got two rejections the AI hadn't flagged: a mandatory screen you couldn't get past in a small window, then a rating request shown too early. Both are told in why Apple rejects an app.

A secret key inside the app. When the AI asks you for an API key, it happily puts it in the code shipped to users, where anyone can read it. The guide to API keys in an Expo app explains where to make the call.

The fix is the same every time: one check per risk. Access rules tested with two accounts, a backup before every database change, a test on the smallest and largest iPhone, no secret key in the app.

A first project: a checklist for an outing

Let's take a fictional example: an app that helps you pack your hiking bag. The first version only lets you add an item, check it off and keep the list.

Start with a prompt that describes a result you can check:

I'm building a small mobile app to pack a hiking bag.
First step: a single screen with a list of items.
I want to add an item and check it off when it's in the bag.
Use local data first, with no user account.
Before changing the project, explain which files are involved.
After the change, tell me how to check adding and checking off.
Don't add payments, maps or AI features for now.

This prompt is an example to adapt, not an app that's already built. It keeps you from handing ten features to the tool when you don't yet know how to check the first one.

Once the screen shows up, test one action at a time. Add an item, check it off, then close and reopen the app. If the list disappears, ask specifically for the data to persist across restarts. If some text is cut off, share a screenshot and mention the device size.

The checks to keep from day one

For every small step, answer these questions:

1. Does the requested action actually work?
2. Is the result still correct after closing the app or going back?
3. Are the keyboard, text and buttons still usable on my phone?
4. Can I go back to the previous version if the next change breaks something?

When the AI suggests a dependency or a service, ask why it's needed and what it costs.

Can you publish a vibe-coded app?

Yes. Solys and TradaZen are on the App Store, and their code was written with Claude and Cursor. On TradaZen, 26 calendar days passed between the first commit and publication.

But how you build it doesn't exempt you from anything. You have to test the app, prepare its listing and explain how it works to Apple. A preview that works in the browser doesn't prove the build sent to Apple works on iPhone. And four steps stay entirely on you: app signing, TestFlight, subscriptions and review.

The App Store publishing guide covers those steps. To get started with OpenAI's tool, use the tutorial for building an app with Codex. With Anthropic's, follow building an app with Claude Code.

Is vibe coding free?

To try it, yes. Cursor, Lovable and Bolt have a free plan, limited in number of requests per day. It's enough for a few screens, not for building an entire app: the quotas run out fast once you work every day. Claude Code, on the other hand, requires a paid plan, starting at $20 per month.

Then come the costs AI doesn't remove: $99 per year to publish on the App Store, $25 one time for Google Play. The breakdown of what's free and what isn't is in building an app for free.

How do you learn vibe coding?

On your own, you can go far: the official docs are free, and the tutorials on this blog start from a real project. Count on practice above all. One small finished app teaches you more than ten videos.

What's missing when you learn alone is knowing, before you submit, what Apple will reject. That's what the iOS Kit is for: the starter code from my apps, a four-step path to follow at your own pace, and Claude Code skills that review your project before submission. The vibe coding course page covers it for AI in general, and the Claude Code course for Anthropic's tool.

Frequently asked questions

Do you need to know how to code to vibe code?

No, not to get started. You describe what you want and the AI writes the code. To finish an app, though, you have to learn to verify: see which files changed, test on your phone, understand an error message.

What is a vibe coder?

Someone who builds software by describing it to an AI rather than writing the code. The word covers both a beginner launching their first prototype and a developer who delegates the writing to go faster.

Can you vibe code with ChatGPT?

Yes. You can ask it for code and paste it into your project, or use Codex, OpenAI's agent, which edits the files directly. The limit is the same as with other tools: you're the one checking the result on a real phone.

What's the best vibe coding tool?

The one that produces the format you need. For a website or web app, a generator like Lovable or Bolt is fast. For an iPhone app you want to publish and keep improving, an agent that works in your own project, like Claude Code, Codex or Cursor, leaves you the code and the control.

Can you build a mobile app with vibe coding?

Yes, as long as you pick a tool that produces a real mobile project and not just a web app. With React Native and Expo, a single project gives you both the iPhone app and the Android app.

Is vibe coding reliable for a production app?

Not as is. The known incidents all come from results published without checks: exposed data, a deleted database, Apple rejections. With tested access rules, backups and on-device testing, the same method produces apps that hold up.

From first idea to a project you can finish

If you're still unsure about the first feature, start by validating your app idea. To see the whole path, read how to build an app. If you already have a prototype, the four-step path in the iOS Kit helps you figure out the next one. Not sure where you stand? The free diagnostic tells you where you are and which step to tackle first, and sends you my Apple anti-rejection checklist.