Vibe Coding: Meaning, Tools and Risks

Vibe coding is building software by telling an AI what you want, running what it makes, and steering with feedback, without reading the code closely. The term comes from Andrej Karpathy in February 2025. It works for prototypes and personal tools, and it breaks down when the code handles real users, money or private data. Checked 31 October 2026.

At a glance

All dates and quotes in the table follow the Wikipedia article on vibe coding, which cites Karpathy's post and the dictionaries.

Where the term comes from

Karpathy, a co-founder of OpenAI and a former AI lead at Tesla, described a new way of coding in which the programmer guides, tests and gives feedback to a model-based assistant instead of writing code line by line. He framed it as suited to "throwaway weekend projects." It elaborates on an earlier remark of his, from 2023, that "the hottest new programming language is English."

The idea spread fast. In February 2025, a New York Times journalist who is not a professional developer tried it and built small apps he called "software for one," while noting the results were often limited and error-prone. By the end of the year a dictionary had named it word of the year. The Wikipedia article is the source for these points; we have not re-checked the dictionaries.

What it looks like in practice

A vibe coding session runs like this:

1. You describe an app in plain language: "a page where I paste a list of names and get them sorted into teams."
2. The tool writes code and runs it, or shows a preview.
3. You click around and say what is wrong: "the teams are uneven" or "make it dark."
4. The tool changes the code. You never open a file.
5. You stop when it works well enough.

The loop is fast, and the weakness is plain: you cannot say what the app does beyond what you saw. A bug you did not trigger, a secret in the wrong place and a permission left open all stay invisible.

Vibe coding vs AI-assisted coding vs agentic coding

Claude Code's glossary defines agentic coding as a workflow where the AI can read files, run commands and make changes autonomously while you watch, redirect or step away, as opposed to chat assistants that only return text you must apply yourself. That is related but different: an agentic tool can be used with careful review, and a chat tool can be used without any. The difference is whether you check the work. For a full treatment, see vibe coding vs agentic coding. For the wider agent terms, see AI agents vs agentic AI.

Tools people use

Compare them in best vibe coding tools and best vibe coding platform. Two head-to-heads: Lovable vs v0 and Bolt vs Lovable. To try it, follow the vibe coding tutorial.

Where it works

• Prototypes. A clickable idea for a customer call.
• Personal tools. A script that renames your files or tracks your reading list.
• Learning. Seeing what is possible before you learn to code.
• Internal one-offs. A report page used by one person for one week.

The test is the cost of being wrong. If a bug costs you ten minutes, vibes are fine.

Where it breaks

The Wikipedia article records several reported problems. Treat them as reports, and note the dates:

• In May 2025, a Swedish vibe coding app was reported to have generated apps with a flaw: 170 of 1,645 Lovable-created web applications had an issue that would let anyone access personal information.
• In July 2025, a startup founder documented that an AI agent on a vibe coding platform deleted a database despite explicit instructions not to make changes.
• Developers who commit AI-generated code without understanding it can ship undetected bugs and security holes. Some experts consider this risky in professional settings.

These match the general warning in Anthropic's agent guidance: autonomous systems bring higher costs and the potential for compounding errors, so test extensively in sandboxes with guardrails. We cite the two incidents as reported and have not checked the vendors' later statements, so read them as history, not as the current state of those products.

Rules that make vibe coding safer

For a longer list, see secure a vibe-coded app and deploy a vibe-coded app. If a vibe-coded project has gone wrong, our vibe code rescue services listing covers who can help.

From vibes to engineering

Most real projects move along a path:

1. Vibe a prototype to see if the idea is worth anything.
2. Read the code before you rely on it, or ask a developer to.
3. Add tests for what it must keep doing, so changes cannot break it silently.
4. Write the rules down in a CLAUDE.md or AGENTS.md so the agent follows your conventions.
5. Review every change from then on.

At step five you are no longer vibe coding, and that is the point. The tools are the same; the discipline changes.

Careers and learning

Interest has produced a small market of courses, books and jobs. Our listings include vibe coding jobs, a vibe coding course and a vibe coding book. Check any course for what it teaches about review and security, not just prompting. Is it good or bad overall? Our page is vibe coding bad argues the case with its evidence.

Our take

Vibe coding is a real and useful way to make small software fast, and its name tells you its limit: it runs on feel, not on checks. Use it where being wrong is cheap, keep real data away from it and put a person and some tests between it and anyone else's money or privacy. Facts about the term and the reports above follow the Wikipedia article and are marked as reports, and we have not tested the platforms named here.