What is vibe coding? The honest version

It’s 1:40 in the morning and someone who has never written a for-loop is arguing with a computer about a button. “Make it less shouty.” The computer makes it less shouty, and in the process breaks the login page. They paste the error back in. It fixes the login page. By 3am there is a working app that tracks which houseplants they’ve forgotten to water, and they could not tell you a single thing about how it works.

That’s vibe coding. Not all of it, but the heart of it: you describe what you want, an AI writes the code, and you steer by looking at the result rather than reading the code.

Where the word came from

In February 2025, Andrej Karpathy (a founding member of OpenAI and formerly head of AI at Tesla) posted about a new way he’d been building throwaway projects. He described what he wanted, accepted whatever the AI wrote, pasted error messages back in without comment, and kept going until it worked. He called it vibe coding, and suggested you could “forget that the code even exists.”

The name stuck because it described something a lot of people were already doing and felt faintly guilty about. By the end of the year Collins Dictionary had made it their word of the year, and people had started putting “vibe coder” in their bios with a straight face.

What it actually looks like

There’s no membership test. Here are three people who would all fairly call themselves vibe coders:

• The one who has never opened the code. Uses Lovable, Bolt or v0. Describes the app in plain English, clicks around the preview, describes what’s wrong. Has shipped real things. Has never seen a terminal and would like to keep it that way.
• The developer with a very fast intern. Works in Cursor or Claude Code. Writes a plan, lets the agent write most of the code, skims the changes, and steps in when it does something daft. Probably doesn’t say “vibe” in front of their manager.
• The in-betweener. Started in the first group, got curious the third time something broke, and now knows just enough to be dangerous. A lot of people on Vibers live here, and it’s a great place to be.

Purists will tell you only the first one counts, because Karpathy’s original point was not looking at the code at all. In practice nobody checks credentials at the door. If an AI wrote most of it and you did the steering, you vibe coded it.

The loop

Whatever the tool, every session runs on the same four steps: describe, look, complain, repeat.

The skill is almost entirely in step three. “It’s broken” gets you a coin toss. A precise complaint gets you a fix.

the dashboard is broken fix it

On the dashboard, new users see an empty chart because they have no data yet. Replace the empty chart with a friendly message and a button that says “Add your first entry”. Don’t change anything else on the page.

That last sentence is doing a lot of work. AI tools are eager to help, and left alone they will “improve” three things you didn’t ask about while fixing the one you did.

What it’s brilliant at

• Idea to clickable thing in an afternoon. The distance between “wouldn’t it be nice if” and a working version has collapsed.
• Software for an audience of one, or ten. The wedding countdown, the shop’s stock tracker, the tool that formats your weekly report. Things no company would ever build for you.
• Testing whether anyone cares. Build the rough version, show it to people, and find out before spending real money.
• Front ends, dashboards, internal tools and small games. The visual, self-contained stuff is where AI is strongest.
• Learning. Asking “why did you do it that way?” turns out to be the most patient programming tutor ever made. It’s awake at 3am and it never sighs.

Where it falls over

Most vibe coders hit what people call the 80% wall. The first 80% of a project feels like magic. The last 20% feels like wading through porridge. The usual culprits:

• Logins and permissions. It’s easy to make a login screen that looks right. It’s much harder to make sure user A can never see user B’s data.
• Money. Payments, refunds and subscriptions are the bits where a bug costs actual cash.
• Secrets. AI tools will cheerfully put a private API key somewhere public if you let them.
• Growth. Each new feature gets bolted onto the last. Eventually the AI loses track of its own codebase, and so do you.

None of this means don’t. It means know where the trapdoors are.

Is it “real” programming?

This argument comes round every few months and generates a lot of heat. It’s a bit like asking whether photography is real painting. Different craft, different skills, plenty of overlap.

What vibe coding definitely is: real making. Things exist that didn’t exist before. People use them. And the skills that make someone good at it (clear thinking, taste, knowing what to build and when to stop) are the same skills that always separated great engineers from merely fast ones. The typing was never the hard part.

A short glossary

Word	What it means
Prompt	What you type to the AI. A good one says what, why and what not to touch.
Agent	An AI that doesn’t just answer but takes actions: edits files, runs commands, checks its own work.
Context	Everything the AI can “see” at once: your messages, your files, the errors. When it forgets things, it’s usually run out.
Hallucination	When the AI confidently uses something that doesn’t exist, like a made-up function or setting.
Diff	The list of changes the AI made. Skimming it is the cheapest safety check there is.
Repo	Your project’s folder of code with its full history, usually on GitHub. Your escape hatch if a tool lets you down.
Deploy	Putting it on the internet so other people can use it.

How to start

Pick something embarrassingly small, ideally a problem you had this week, and build it this weekend. We wrote a step-by-step plan for exactly that: how to vibe code your first app. If you’re stuck on which tool to use, the best vibe coding tools ranks them by what people actually ship.