Vibe coding: what it actually means (and how not to mess it up)

You have an idea for a project, but you don’t really know how to code, or only a little. Over the past year, a word has spread to describe what you’re doing when you describe your idea to an AI and let it write the code for you: vibe coding. Let’s look at where the word actually comes from, what it really lets you do, which tools fall under it, and how to use it without ending up with something unusable or unsafe.

Where the word actually comes from

On February 2, 2025, AI researcher Andrej Karpathy (an OpenAI co-founder and former head of AI at Tesla) posted a message that would go on to become famous: “There’s a new kind of coding I call ‘vibe coding’, where you fully give in to the vibes, embrace exponentials, and forget that the code even exists.” He described his own practice: talking out loud to his code editor, accepting almost every suggestion the AI made, and when something broke, copy-pasting the error back in and letting the AI fix it. His own summary said it all: “I just see stuff, say stuff, run stuff, and copy paste stuff, and it mostly works.”

Karpathy later clarified that he originally meant small, throwaway weekend projects, not software meant to last. The word escaped that framing anyway: in November 2025, Collins Dictionary named it their Word of the Year, with this official definition: “the use of artificial intelligence prompted by natural language to write computer code.”

What it lets you do when you’re not a developer

In practice, vibe coding flips the usual logic of programming. Instead of learning a syntax and writing every line yourself, you describe what you want in plain words: “build me a site so members of my student club can sign up for time slots,” “add a button that exports this list as a PDF.” The AI proposes the code, runs it, shows you the result. You react, correct, refine, and the loop continues.

For a student outside computer science, that opens a real door: testing an idea, building a small tool for a thesis, a student club, an association, or a personal project, without spending six months learning a language before seeing anything work. It isn’t magic, though: you’re still the one deciding what to build and in what order, and still the one checking that the result matches what you actually wanted. The AI writes the code; it doesn’t replace your judgment about what should exist in the first place.

The tools that fall under this word

The term covers several families of tools, aimed at different people:

• Claude Code: a command-line tool from Anthropic. It reads your project, proposes changes, and runs commands. Built with developers in mind, but usable by anyone willing to open a terminal.
• Cursor: a code editor with AI built in, better suited to someone who already reads code and wants to move faster on an existing project.
• Lovable: built explicitly for people who don’t code, with a visual interface: you describe your app in one instruction and get a full project back.
• Bolt.new: built by StackBlitz, it builds a complete application directly in your browser from a prompt.
• v0: Vercel’s tool, more specialized in generating interfaces (screens, components) than a full application with its own database.
• Replit: a browser-based development environment, with an agent that can build, test, and fix an application on its own.

Each has its own strengths and its own audience. None of them excuses you from thinking about what you’re building before you start.

The limits you need to know before diving in

The first limit comes from Karpathy himself: he was talking about throwaway projects, not code meant to last or to hold real users’ data. Used without a method, vibe coding quickly produces something that looks like it works on screen while hiding real gaps underneath.

This isn’t theoretical. In late September 2026, an investigation found around 16,000 Supabase databases (a database service heavily used by these AI tools) leaking personal data such as names, addresses, or passwords, because of misconfigured applications. The dedicated article on this blog about securing a vibe-coded app walks through the exact mistakes to fix before you publish anything.

The second limit is quieter: the AI does what you ask, not necessarily what you meant. A vague request produces a vague result. That’s exactly why Anthropic itself, which builds Claude Code, recommends separating exploration and planning from writing code, rather than diving straight in: let the AI read the project and propose a plan before letting it change anything.

A 5-step method so you don’t mess it up

1. Write down what you want first, on a single page, before you even open an AI tool: the problem you’re solving, who will use it, the must-have features. The next article on this blog gives you a ready-to-copy template.
2. Ask for a plan before any code. Describe your project, then ask the AI to propose an approach and list the files involved, before letting it change anything. A simple prompt does the job:

Here's my project: [describe your idea in 2-3 sentences].
Before changing anything, explore what already exists
and propose a step-by-step plan. Don't write any code
until I've approved that plan.

3. Move in small steps. One feature at a time, reviewing what changed before moving to the next.
4. Check before you publish. An app that “works” on screen can still be leaking data or secret keys. That deserves its own article, further along on this blog.
5. Keep a record of every version, with a tool like Git for instance, so you can roll back if a step breaks everything.

If the project is graded coursework, check your school’s AI usage rules first: they vary a lot between institutions, and some strictly regulate these tools for graded submissions.

Key takeaways

• Vibe coding, popularized by Andrej Karpathy on February 2, 2025, and named 2025 Word of the Year by Collins Dictionary, means building software by describing what you want rather than writing every line yourself.
• Several tools fall under the term, for different audiences: from the terminal (Claude Code, Cursor) to visual interfaces built for non-developers (Lovable, Bolt.new, v0, Replit).
• Karpathy himself reserved it for throwaway projects: without a method, a more serious project can hide real security gaps.
• A good method always starts before the AI does: a mini spec, an approved plan, small steps, a record of every version.
• For graded work, check your institution’s rules before using these tools.

To write your mini spec before you start, you can work from the template in the cahier-des-charges-vibe skill.