What is vibe coding, and when does it break?

Vibe coding means building software by describing what you want to an AI and accepting the code it writes without reading it closely. You judge the result by running it: if it works, you keep going, and if it breaks, you paste the error back in. Andrej Karpathy named it in February 2025, and Collins Dictionary made it its word of the year.

It is a great way to make prototypes and personal tools. It breaks when real users, real data or real money are involved, because nobody has checked what the code actually does. This guide covers where the term came from, what it is good for, and how to vibe code without getting burned.

01Where the term came from

On February 2, 2025, Andrej Karpathy, a founding member of OpenAI and former head of AI at Tesla, described a new way of coding on X. You go with the vibes and “forget that the code even exists.” He accepted every change without reading the diffs, pasted error messages back in without comment, and asked for random changes until bugs went away.

He was clear about the limits: he saw it as fine for weekend projects you plan to throw away. The name spread anyway, because it described something many people were already doing.

1. February 2025Karpathy coins “vibe coding” in a post on X
2. March 2025Simon Willison defines it narrowly: building software with AI without reviewing the code it writes
3. November 2025Collins Dictionary names “vibe coding” its word of the year
4. February 2026Karpathy suggests “agentic engineering” for the careful, professional version

Collins defines vibe coding as the use of AI, prompted by natural language, to write computer code. That broad sense is how many people use it today. The narrow sense matters more, though, and developer Simon Willison put it best: if you reviewed and tested the code and can explain how it works, that is not vibe coding. It is software development, whoever typed it.

02What vibe coding looks like in practice

A vibe coding session is a loop. You describe, run, react and repeat, and you never open the code unless something forces you to.

Build a single-page web app where I can paste a list of expenses and see a pie chart by category.
Save everything in the browser so it is still there when I come back.
Make it clean and simple, with a dark mode.

Then the follow-ups are reactions, not instructions: “the chart is empty,” “make the numbers bigger,” or a pasted error message. Within an hour you often have something that works well enough to use.

People do this in chat-based app builders like Lovable, Bolt, v0 and Replit, and in coding agents like Claude Code, Codex and Cursor. Our comparison of the best AI coding agents of 2026 covers the differences. Any of these tools can be used carefully or carelessly. What makes it vibe coding is the lack of review, not the tool.

Most developers kept vibe coding out of their day jobs, at least in 2025. In Stack Overflow’s survey that year, 72% of respondents said it was not part of their professional work, and another 5% said so emphatically. About 15% said it was, to some degree.

03What is vibe coding good for?

For low-stakes work, vibe coding is hard to beat. Even Willison, who is strict about AI code in production, says to go for it on low-stakes projects and prototypes.

The pattern is simple. The cost of a hidden bug decides whether vibe coding is fine. If the worst case is “I rebuild it on Saturday,” go ahead. If the worst case is “a stranger reads my users’ messages,” slow down.

04Where vibe coding breaks

It breaks in three predictable places: security, maintenance and scale.

Security. AI writes code that looks right far more often than it writes code that is safe. Veracode has tested more than 150 models, and two years of new releases have barely moved its security pass rate.

A real example: in January 2026, security firm Wiz looked at Moltbook, a social network for AI agents whose founder said he did not write a line of its code. A database key sat in the site’s public JavaScript, and row-level security (the database rules that limit who can see which rows) was off.

Anyone could read and change the production database. Wiz counted 1.5 million API tokens, 35,000 email addresses and thousands of private messages. The team fixed it within about three hours of being told. It is exactly the kind of mistake a quick review catches.

Maintenance. Karpathy himself admitted that the code grew past what he could easily follow. When nobody understands a codebase, every fix is a guess, and each guess can break something else. Bugs that keep coming back are the usual first sign.

Scale. What works for five users can fall over at five thousand. Slow database queries, missing rate limits and paid APIs called in a loop rarely show up while you click around alone. They show up as an outage or a surprise bill.

05How to vibe code responsibly

You do not have to give up the speed. Add a few guardrails that take minutes, not days.

1. 01

   Start every project in git

   Commit after each step that works. When the AI breaks something, you can roll back in seconds instead of asking it to undo its own changes. Our guide to git for beginners covers eight commands worth knowing.

2. 02

   Keep secrets out of the code

   API keys and database passwords belong in environment variables on the server, never in files a browser can load. Ask the AI where each key lives and who can see it.

3. 03

   Ask for tests as you go

   After each feature, ask the agent to write a test, run it and show you that it passes. Tests catch the regressions you will not notice by clicking around.

4. 04

   Read the risky parts, skip the rest

   Skim styling and layout if you like. Read anything that touches logins, payments, user data or deletion. Our guide on how to review AI-generated code shows how to review by risk.

5. 05

   Run a security pass before anyone else uses it

   Use the prompt below, then fix what it finds. For the full list, see our checklist for securing AI-generated code.

Before I share this app, act as a strict security reviewer.
1. List every place user data is stored or sent, and who can read it.
2. Find any API key, token or password in a file the browser can load.
3. Check that the database only lets each user read and change their own rows.
4. Check every form field and URL parameter for injection and cross-site scripting.
5. List what would break or cost money if 1,000 people used it at once.
Do not fix anything yet. Give me the list, most serious first.

06When to graduate to a careful agent workflow

A year after coining the term, Karpathy wrote that programming through AI agents is becoming the default for professionals, but with more oversight and checking. His favorite name for that is agentic engineering. Agents still write most of the code, but you plan it, check it and own the result.

Switch when any of these become true: other people use it, it stores personal data or payments, you will maintain it for more than a few weeks, or the same bugs keep returning. The AI pair programming workflow is a good next step, and it keeps most of the speed.

07FAQ

Read next: how to review AI-generated code without reading every line, or compare the best AI coding agents of 2026.