What is vibe coding?
By Mari Luukkainen · 7 min read · Last updated
On this page
- What did Karpathy mean by vibe coding?
- How does a vibe coding session work?
- What can you build with vibe coding?
- Which tools do people use?
- What has gone wrong with vibe coding in public?
- How is vibe coding different from no-code and from normal programming?
- What does the first 30 minutes look like?
- How I started
- Frequently asked questions
Vibe coding is building software by describing what you want to an AI in plain language, running the result and telling the AI what to change. Andrej Karpathy coined the term in February 2025. You direct, test and decide, and the AI writes most of the code.
This page covers the original definition, how a session works with a real prompt, what you can build, which tools exist, what has gone wrong in public and how I started. You need no coding experience to follow it. Unfamiliar words are in the glossary.
Ask AI about this pageChatGPTClaudePerplexityGoogle AI ModeGrok
What did Karpathy mean by vibe coding?
Karpathy posted the term on 2 February 2025 (original post). His description: "you fully give in to the vibes, embrace exponentials, and forget that the code even exists." He said he asks for small changes by voice, clicks "Accept All" every time, no longer reads the diffs, and pastes error messages back in without comment. He called it "not too bad for throwaway weekend projects."
The definition that matters in practice comes from Simon Willison, written in March 2025: vibe coding is "building software with an LLM without reviewing the code it writes." He adds that it is "not the same thing as writing code with the help of LLMs" (post). If you read and understand every change, you are doing AI-assisted programming. If you accept changes and judge the result only by running it, you are vibe coding.
Collins Dictionary named "vibe coding" its word of the year for 2025 and defined it as "the use of artificial intelligence prompted by natural language to assist with the writing of computer code" (CNN, 6 November 2025). That definition is wider than Karpathy's. Most people now use the word for both styles.
The line I use: on a prototype or a personal tool I accept changes without reading them. On anything that holds other people's data I want a review. Is vibe coding bad covers where that line sits.
How does a vibe coding session work?
- Pick an idea that fits in two sentences.
- Describe it: who uses it, what happens on each screen, what the main button does and what is out of scope.
- The AI builds a first version. A browser builder shows a preview. An editor or terminal tool shows files appearing in a project folder.
- Use it like a stranger would: click everything, leave fields empty, open it on a phone.
- Report what is wrong in one sentence, or paste the error message.
- The AI fixes it. Repeat steps 4 to 6 in small rounds.
- When it works, save it with Git and publish it.
Step 5 in practice: the tester finds that a recipe scaler shows NaN when the amount is typed with a comma. The prompt has four parts: what you did, what you saw, what you expected, and a limit on what may change.
Bug: when I type 1,5 as an amount (with a comma), the scaled amount shows NaN.
Expected: treat the comma as a decimal point, so 1,5 is 1.5.
Only change the amount parsing. Tell me which file you changed.The full walk-through with a first project is how to vibe code.
What can you build with vibe coding?
| Project | Example | Difficulty for a first project |
|---|---|---|
| Landing page | A page for a bakery with opening hours and a map | Easy |
| Personal tool | A recipe scaler, a habit tracker, a budget sheet | Easy |
| Internal tool | A form that fills a spreadsheet, a team dashboard | Medium |
| Prototype | A clickable version of a product idea to test with five people | Medium |
| Automation script | Rename 400 files, collect prices from a page | Medium |
| Product with accounts and payments | A paid app with users | Hard: needs review by someone who reads code |
My own range: in July 2025 I built a pitch deck tool, Pitchkit.net, in six days, and it made 700 EUR of monthly recurring revenue the following week. I have also built the deploy platform and the monitoring system that run my 100+ apps (stuff I built to manage 100+ apps).
Which tools do people use?
- Browser builders. Lovable, Bolt.new, Replit, v0, Base44 and Bubble. You chat and see a preview, with nothing to install. See best AI app builders.
- AI editors. Cursor and VS Code with GitHub Copilot. You see the files and each change as it appears. See Cursor vs Claude Code.
- Terminal agents. Claude Code, Codex, OpenCode and Antigravity CLI (the successor to Gemini CLI for individuals). You give a task in a terminal and review the result. See what Claude Code is.
The tool overview has a table for choosing and the current free tiers and prices.
What has gone wrong with vibe coding in public?
Two incidents show the risk better than a general warning does.
- Replit and SaaStr, July 2025. During a test by SaaStr founder Jason Lemkin, Replit's AI agent deleted a production database holding records on 1,200+ executives and 1,190+ companies. The agent told him a rollback would not work; Lemkin recovered the data. Replit's CEO called it "unacceptable and should never be possible" (Fortune, 23 July 2025).
- Lovable, CVE-2025-48757. Published 30 May 2025 with a CVSS score of 9.3. The entry says an insufficient database Row-Level Security policy in Lovable through 15 April 2025 let unauthenticated attackers read or write arbitrary database tables of generated sites. Lovable disputes the classification and says each customer is responsible for their own project's configuration (NVD).
Both come from the same gap: the person accepted what the tool built and did not check what it could reach. The checks are in is vibe coding bad.
How is vibe coding different from no-code and from normal programming?
| Approach | You write | You can change | Typical limit |
|---|---|---|---|
| Traditional programming | Code, line by line | Anything | Needs years to learn |
| No-code tools | Visual rules and settings | What the tool offers | The tool's feature list |
| Vibe coding | Descriptions and corrections | Anything the AI can write and you can test | Your ability to check the result |
No-code tools limit what you can build. Vibe coding limits what you can verify, which is a different problem: the app can do anything, and you still have to know whether it does it correctly.
What does the first 30 minutes look like?
| Minutes | What happens |
|---|---|
| 0 to 5 | Write the idea in four answers: who uses it, what each screen does, what the button does, what is out of scope |
| 5 to 10 | Paste it into a tool and read the plan it proposes. Correct one wrong assumption |
| 10 to 15 | The first version appears. Expect it to look good and have a couple of bugs |
| 15 to 25 | Test with odd input and send fix prompts, one at a time |
| 25 to 30 | Commit the working state and ask the tool to explain the files |
Publishing takes a further 15 to 30 minutes the first time. This timing is an estimate for you to check against your own run, not a measurement.
How I started
In March 2025 I built my first Next.js app on a two hour flight from Helsinki to Berlin. I was a VC and kept seeing vibe coded products from non-technical founders in our dealflow. I asked ChatGPT what this was, and it pointed me at Cursor. Fifteen years earlier I had built LAMP applications and had never used GitHub. On that flight I learned GitHub and the modern stack I had skipped (first-year post).
The first attempts went badly. I was asking ChatGPT "What's md for Mac?" and breaking things with typos. It started working when I planned in ChatGPT first, broke the work into pieces and only then touched code (post).
I quit VC three months later and have vibe coded full time since July 2025. I have pushed more than 10,000 commits to GitHub and built over 100 apps. These days I barely read the code. What makes that workable is a private handbook of decisions that the AI reads before it starts, and guardrails I added after specific outages. By Willison's definition, much of my work is vibe coding, and the handbook is how I limit the damage.
Frequently asked questions
Is vibe coding the same as AI-assisted programming?
No. Simon Willison's distinction is whether you review the code. If you read and understand each change, it is AI-assisted programming. If you accept changes without reviewing and judge only by running the result, it is vibe coding.
Was vibe coding really the Collins word of the year?
Yes. Collins Dictionary named it word of the year for 2025, announced on 6 November 2025.
Do I need to know how to code to start vibe coding?
No, a first project needs no coding knowledge. Learning to read code later helps you judge results and fix problems the AI cannot solve alone.
How long does it take to vibe code a first app?
A small tool such as a recipe scaler takes an evening, including publishing. My first Next.js app was built during a two hour flight. A product with accounts and payments takes longer and needs a review.
Can vibe coded apps be used for real businesses?
Yes. My Pitchkit.net, built in six days, made 700 EUR of monthly recurring revenue the week after launch. Before you take customers' data or payments, have someone who can read code review it.
Next guideHow to vibe code →