Where the term came from.
Andrej Karpathy — formerly of OpenAI, formerly of Tesla — coined the phrase on Twitter in early 2025. He described a way of programming where you barely look at the code: you describe what you want, accept what the AI generates, run it, see what happens, describe a fix, repeat. He called it “giving in to the vibes,” hence the name.
The phrase caught on because it named something a lot of people had quietly started doing — including engineers who knew better, including people who’d never written a line of code in their life. Within a few months, “vibe coding” was shorthand for an entirely new shape of work.
What’s actually happening.
Three things had to land at the same time for vibe coding to become possible:
Models got good enough to be trusted with execution. By 2025, Claude and GPT could generate working code on the first try for a wide enough range of tasks that you could stop second-guessing every line. Not always. Often.
Tool-calling matured. Models stopped being chatty text-generators and started being agents — capable of running scripts, reading files, calling APIs, browsing the web. That collapsed the loop between “describe” and “running thing.”
Coding harnesses showed up. Tools like Cursor, Claude Code, and the IDE plugins built around them turned the model into a collaborator that could see your whole project, edit multiple files at once, and remember what it was working on. Suddenly the model wasn’t just writing snippets — it was navigating real codebases.
What it is not.
Vibe coding isn’t low-code or no-code. Those are tools that replace code with a different abstraction (drag-and-drop, spreadsheets, configuration). Vibe coding still produces real code in a real language. You just write a lot less of it yourself.
It also isn’t pair programming with an AI assistant in the old sense — the GitHub Copilot mode of “suggest the next line.” Vibe coding is closer to directing than writing. You describe outcomes, the AI handles the mechanics, and you read what comes back to confirm it’s doing what you actually meant.
And it’s not magic. The model still gets things wrong. It still has knowledge cutoffs, it still hallucinates APIs that don’t exist, it still produces code that compiles but does the wrong thing. The skill is knowing how to spot that, fast, and steer.
Who’s actually doing this.
Three groups, mostly:
Non-engineers shipping their first real software. Founders prototyping, designers building tools, operators automating their own jobs. People who could describe what they wanted but couldn’t implement it before — and now can.
Senior engineers operating at 5–10×. The counterintuitive group. The engineers who already know what they’re doing — who can code fluently, who don’t need the AI for syntax — get the biggest leverage, because they can judge what the AI produces and steer it precisely.
Hobbyists building things they wouldn’t have otherwise. The category that probably matters most for culture. People building little weather widgets, family cookbooks, neighborhood tools — software that wouldn’t have existed because the activation energy used to be too high.
What you actually do.
On a good day, vibe coding looks like this: you open Cursor or Claude Code, describe the thing you want in a couple of sentences, watch a small wave of edits land across the relevant files, run the result, and either ship it or tell the AI what went wrong. The cycle takes seconds to minutes. You go from “I wonder if” to a working prototype in the same sitting.
On a bad day, you spend twenty minutes chasing a bug the AI confidently introduced and refused to acknowledge — until you stopped, read the code yourself, and pointed it at the line it had hallucinated.
The skill isn’t typing. The skill is describing precisely, noticing when something is wrong, and knowing when to stop trusting the AI and read.
How to start.
The honest fastest way is to do it with someone in the room who can shortcut the failure modes. That’s why VibeFest exists.
Four ways in, depending on how deep you want to go:
- Vibe — one hour · the lowest-stakes version. Find out if this is for you.
- Learn — two hours · one technique, one workflow you actually use Monday.
- Build — a full day · walk out with a working AI agent on your own Mac.
- Ship — five days, for two · co-build a system you can deploy immediately, with the partner of your choosing. Two Macs to take home.
Or read the rest of the way in: VibeFest — the whole picture.