What is vibe coding, and can you make games with it?

What is vibe coding?

Vibe coding is building a program by talking to AI in everyday language and accepting what it writes without reading the code. You say what you want, the AI writes it, you see if it worked and, if there is an error, you paste the error message back into the chat.

"Vibe" here means mood, feeling. The idea is to code by feel: you steer the work by the result on the screen, not by the code behind it.

For people who have never programmed, this opened a door that did not exist before. For people who do program, it became a debate: how far can you trust code nobody has read?

Where did the term vibe coding come from?

From a post by Andrej Karpathy, a cofounder of OpenAI and former head of AI at Tesla, published on X on February 2, 2025¹. He described a new way of programming where you give in to the vibes and forget the code even exists, and said it was not bad for throwaway weekend projects².

The name caught on fast. By March 2025, Merriam-Webster was already listing "vibe coding" among trending slang¹. On November 6, 2025, Collins Dictionary chose the term as its word of the year. The Collins definition is the use of artificial intelligence, prompted by natural language, to help write computer code³.

Notice one detail: the person who coined the term was talking about throwaway projects. The idea started as a weekend toy, not as a way to build products.

Is vibe coding the same as programming with AI?

Not exactly, and the difference matters. Simon Willison, who writes about programming with AI on his blog, proposed a simple test: vibe coding is building software with AI without reviewing the code it writes².

If you review it, test it, and can explain what the program does, Willison says that is normal software development, just with AI help². For him, vibe coding works when the stakes are low: a small project, no other people's data, no exposed passwords, and nothing that could run up an uncontrolled bill. Anything other people will use deserves a check from someone more experienced first².

The Collins definition is broader and covers any use of AI to write code³. In practice, both meanings coexist: when someone says "I vibe coded this," they could mean a game made in an afternoon or a months-long project reviewed with care.

What is it like to make a game with vibe coding?

It is a back-and-forth conversation. You describe the game, the AI builds a first version, you play it, point out what is wrong, and it fixes it. When an error shows up, you copy the message and hand it to the AI to solve.

Ethan Mollick, who writes about AI in the newsletter One Useful Thing, described in March 2025 what it was like to make a game this way with Claude Code. He started with a single request: a 3D game where he placed buildings in a city and drove around it. Then he asked for a fire truck, traffic, spreading fire, day and night, and missions⁶.

The AI usage cost about 5 dollars. On the third request, the game stopped working. Mollick did not know the language the game was written in, so he kept passing the errors back to the AI, a back-and-forth of about twenty minutes that cost another 8 dollars or so. The problem turned out to be simple⁶.

His conclusion is the most useful part: vibe coding does not eliminate knowledge, it moves it. Instead of writing the code, you need to know enough to guide the work, find the problem, and judge the result. And the tools, in his view, are not yet reliable enough to work on their own⁶.

Can you make games with vibe coding?

Yes, and there are already plenty. The better question is what kind of game comes out this way, and how much of it can survive going into a store.

• A competition for AI-made games. Vibe Jam 2026, organized by Pieter Levels, required at least 90% of the code to be written by AI and received 945 games between April 1 and May 1, 2026. The 2025 edition passed a thousand entries in its first month, according to the organizer, and had Karpathy among the judges⁵.
• The browser flight game. In March 2025, Pieter Levels, who was already building software and internet businesses and had a large following, said that fly.pieter.com, a free flight simulator played in the browser and made with AI, was bringing in more than 50,000 dollars a month. The number is his own, unaudited, and the headline of the 404 Media story warns that yours probably won't make that⁴.
• The podcast team. In 2026, the host and two producers of a Slate podcast, none of them programmers (one had a little experience), used Claude Code to make a game in the style of The Oregon Trail about making a podcast. They themselves called the game ridiculous, and the story's headline says it turned out bad⁷.
• The accountant who made it to the app stores. Jonathan Bjerk, an accountant in Norway, used AI to make Curve Clash, a multiplayer snake game, and launched it on the web, iPhone, and Android in less than seven months. He did not let the AI decide the game: he designed every mode, tuned every detail, and tested a lot⁸.

None of these cases proves there is money in it. Nobody can promise that a game made this way will earn anything, and most games that launch sell very little. What they show is that making something work has become fast, and that making it good still takes time.

Where does vibe coding break in a game?

It breaks wherever nobody checks. The 2026 tests show how big the problem is when AI works alone:

The authors of the second test describe the pattern: agents get the recognizable mechanic right, but deliver games with little content, visual feedback that does not work, and presentation that does not hang together¹⁰. It is the game that "works" and that nobody wants to play twice.

There are also limits the makers themselves acknowledge. The Claude Code documentation warns that in a long conversation the AI starts to forget earlier instructions and make more mistakes, and that it stops when the work looks done. Without a way to check, the one doing the checking is you, error by error¹¹.

And then there is security. In a Veracode study with 80 programming tasks and more than 100 AI models, the generated code worked, but had a security flaw in about 45% of cases. According to the company, the models got better at writing correct code, not at writing secure code¹². In a single-screen game with no internet connection, this matters little. In a game with accounts, purchases, or online leaderboards, it matters a lot. A concern like that is why the podcast game was kept off Slate's website⁷.

Why does accepting everything unchecked produce a generic game?

Because when you do not decide, the AI fills the gap with whatever is most common. A vague request comes back as the average of every similar game out there.

Japanese artist Kazuki Takakura, who made an action game with AI writing all of the code, says this plainly. For him, direction, fun, style, difficulty, and the feel of the controls depend on physical experience, which AI does not have. Letting AI decide these things tends to produce generic work, made for average taste¹³.

That is why, in his game, the AI handled the programming, the planning, and the bug hunting, while he kept the direction, the art, the sound, and the final say¹³. Bjerk summed up the same arrangement in one sentence: "AI was my engineering team, not my creative director"⁸.

Our take: pure vibe coding, accepting everything, is good for finding out whether an idea is fun. To make a game that feels like yours, someone has to say no to the first result, and that someone is you.

Is vibe coding changing the game stores?

It is increasing the number of games, but not the number of successful ones at the same rate. A South Korean report from July 2026 links the rise in mobile releases to vibe coding: 181,000 games in the six months to May 2026, up 118% on iPhone and 73% on Android in one year, according to ATTN Economy. In the same article, Naavik reports that games earning more than 20,000 dollars grew only 14% between December 2025 and February 2026¹⁴.

On Steam, the largest store for computer games, an independent census of about 53,600 releases shows that by mid-2026 roughly one in three new games discloses AI-made content, almost always art and images¹⁵. Steam asks for this disclosure for what players see or hear; using AI only to write code is not the focus of the rule¹⁷.

There is pushback from toolmakers too. The team behind Godot, a free program for making games, said in June 2026 that it has no AI features and no plans for any, and that low-quality AI-made contributions to the program's code are rejected automatically¹⁶. The rule applies to people helping build Godot, not to people using Godot to make their own games.

The other side

Vibe coding works well for proving an idea, and poorly for delivering a finished game. Karpathy himself was talking about throwaway projects. The 2026 tests show AI on its own getting between 41% and 54% of the work right, the code can come out insecure, and a result with no direction tends toward the generic. If you want to publish, you have to check what the AI does, even without reading code: by playing, testing with other people, and saying no to what is not good enough.

So, is it worth making a game with vibe coding?

As a starting point, yes. A first version made on a whim shows you quickly whether your idea is fun, and throwing it away costs little. To go further, the full path, with the steps and the costs, is in how to make a game with AI.

Your action plan

Check off what you've done. It's saved on this device, so you can pick up where you left off.

1. Write your game idea in three sentences before you open the AI: what the player does, how they win, and why they would play again.
2. Write down one sentence about what makes your game different from similar ones that already exist.
3. Separate what is an experiment, which can go in the trash, from the game you want to publish.
4. Play every version the AI delivers before you accept the next change.
5. Leave logins, purchases, and online leaderboards out of your first game.
6. Ask two or three people to play before you think about a store.