"It's not really about programming; I just see things, say things, run things, copy and paste things, and it generally works."
Can you imagine creating applications in minutes without knowing how to program or write a single line of code? Welcome to the future of software development: this is Vibe Coding.
This is the vision of Andrej Karpathy, former head of AI at Tesla and pioneer of the Vibe Coding phenomenon.
What is Vibe Coding?
It is an emerging approach to artificial intelligence-assisted programming, where the developer describes in natural language what they want the program to do, and an AI model generates the necessary code.
In essence, the programmer’s role changes from writing each line of code to guiding, testing, and refining the code produced by the AI.
The key idea is to “go with the flow” or “unleash the vibe” (hence the name “vibe”) and “forget that code even exists,” delegating the task of “chopping code” to the language model.
For example, Karpathy has created entire projects through dialogue with an AI: he built a Battleship game and a text-reader application without touching the keyboard, dictating commands by voice in about an hour each.
In this conversational paradigm, the developer provides prompts (including voice commands), the AI generates the code, and the human evaluates and adjusts the output. If an error appears, the usual vibe coding practice is to simply paste the error message back to the AI so that it can attempt to correct it automatically.
This allows even people without extensive software engineering training to produce functional applications by simply describing their objectives in natural language.
Tools and platforms that enable Vibe Coding
This development style has gained momentum thanks to various AI tools that facilitate code generation from descriptions. Some of the most prominent platforms for vibe coding are:
- Replit Ghostwriter / Agent: An online IDE with an AI assistant that allows you to build web apps from simple natural language prompts. It guides you step by step, generates the code, and publishes your app in minutes.
- DeepSite: Free in-browser platform for creating websites or games by typing what you want. It shows a real-time preview and generates the code instantly without installing anything.
- Cursor (Composer): Desktop editor with AI that acts as a conversational copilot: modifies functions, generates code and even allows voice programming. Ideal for working on existing code bases.
- GPT-4 / Claude: /Gemini2.5: Language Models (LLMs) like GPT-4 or Claude generate code from descriptions: from simple functions to complete apps. Just say what you want and let the AI build it… Another option is Google’s powerful new multimodal model, Gemini 2.5, which can understand and generate code, making it ideal for prototyping or bug fixing naturally. Available in Bard / Gemini with access to tools.
- GitHub Copilot / Amazon CodeWhisperer: Coding wizards that complete code in real time in your regular editor. They don’t generate entire apps, but automate repetitive tasks and suggest innovative solutions.
Real-life examples of Vibe Coding applications
Numerous developers -from amateurs to professionals- are already adopting vibe coding to create projects in a faster or more accessible way than with traditional methods: I have used it to generate ideas that I had in my head and that saw the light of day in a short time, using this type of tool.
- Non-technical users, such as journalist Kevin Roose, have created useful apps using only natural language, such as LunchBox Buddy, which suggests recipes based on what’s in the fridge. He calls it “software for yourself.
- Independent creators, such as Pieter Levels, developed a multiplayer 3D game in minutes using AI (Cursor + Claude), with no prior knowledge of video game programming. The game was launched and played by more than 22,000 people.
- Startups like Menlo Park Lab are using vibe coding to create products like Brainy Docs, which turns PDFs into explainer videos. They fix bugs by simply re-prompting AI, speeding up development.
- Mass adoption: At Replit, 75% of users no longer write code manually. At Y Combinator, 25% of recent startups generate 95% of their code with AI. The trend is clear: from idea to product with few prompts.
Expert opinion: future of Vibe Coding and impact on job skills
- Sam Altman (CEO of OpenAI) and Mark Zuckerberg (CEO of Meta) anticipate that software engineering will undergo radical changes and that AI will soon perform much of the work typically done by mid-level programmers.
- Harry Law (University of Cambridge) warns that, although it allows apps to be created quickly, vibe coding can lead to reliance on AI without understanding key fundamentals, generating code that is difficult to maintain or insecure.
- Simon Willison (Datasette) qualifies that accurate vibe coding involves accepting code that you don’t fully understand. It is helpful for prototypes, but risky for critical software without human quality control.
- Andrew Chen (a16z) points out that AI can quickly solve the initial 75% of a project, but can complicate the last part with hard-to-tune errors. Frustrating if you are looking for precision.
- Andrej Karpathy describes it as ideal for weekend projects: fun, fast, but not always reliable for complex products.
- Tim O’Reilly (CEO of O’Reilly Media) concludes, “AI won’t replace programmers, it will transform their jobs.” Less hand coding, more design, verification, strategy, and collaboration with AI.
From the classroom to the business: how Vibe Coding is changing who can build
Vibe Coding’s influence goes far beyond how programming is taught. It is a profound cultural shift in the way ideas are transformed into real products.
In the past, many professionals – especially in business, marketing, operations, or even education roles – had powerful ideas, but were stuck in “PowerPoint limbo”: well-intentioned presentations that never became functional solutions because they required a technical team to develop them.
With Vibe Coding, that changes. Today, anyone who can express a need or vision in natural language can start building it with the help of an AI. You no longer need to know how to program to prototype a reservation system, an internal feedback app, an educational game, or an interactive dashboard. Just think like a user, write like a human, and refine like a creator.
This democratizes innovation within organizations:
- A training manager can create their microlearning app.
- An operations manager can design an automated checklist system for their team.
- A salesperson can generate an interactive quote calculator for their customers.
Instead of asking IT to “interpret” their vision, they can now collaborate directly with an AI to build the first working draft. In many cases, what used to take weeks or require a dedicated budget can now be started in an afternoon with just a few lines of well-formulated text.
Moreover, this approach is starting to change how programming is taught: no longer as an isolated technical skill, but as part of the creative and strategic toolkit that any professional should have. Knowing how to converse with an AI to create digital solutions will be as important as knowing how to make a presentation or write a persuasive email….
Delegating code generation to an AI can lead to functional but challenging to maintain, insecure, or poorly structured results if not properly validated. By generating software without a thorough understanding of its inner workings, there is a risk of creating a “black box” that no one on the team can confidently audit or escalate. In addition, bugs that appear to be fixed with a simple re-prompt can hide deeper technical problems, creating technical debt in rapidly expanding projects. While this technique is helpful for prototyping or validating ideas, human oversight remains crucial if the goal is to build robust and sustainable solutions over time.
I hope you found this content interesting 🙂
Have a good week!
Article sources: Karpathy (2025), Business Insider, The Guardian, Ars Technica, TechCrunch, A Computer Scientist on the Evil Side, DeepLearning.AI, among others.
