What is vibe coding?
The word covers a real shift in who can produce working software. It is not a technique you learn so much as a way of working that became possible once models got good enough to hold an entire small application in mind at once.
Vibe coding is describing the software you want in plain language and letting an AI model build it, then steering the result by describing what to change rather than by editing code.
Where the term came from
Andrej Karpathy described the practice in early 2025: giving in to the vibes, talking to the model instead of reading its output line by line, and accepting code you have not fully reviewed because the thing it produces works.
The name stuck because it captured something people were already doing and did not have a word for. It also carried the criticism built in — the reason the phrase is memorable is that "vibes" is not how anyone would describe engineering they trusted.
What it actually looks like in practice
You describe what you want. Something appears. It is wrong in two places, so you say what is wrong, and it changes. You repeat that until the thing does the job.
The loop is conversational rather than technical, and that is the whole point: the skill it rewards is being able to say precisely what you want, which is a different skill from knowing how to build it.
- Describing the outcome, not the implementation, gets better results than trying to specify a solution in prose.
- Being concrete about the data — what a record is, what fields it has — matters more than being concrete about the interface.
- Saying what is wrong with what you got beats restating the original request.
Where it breaks down
The failure mode is not that the model writes bad code. It is that you cannot tell. Vibe coding works when you can evaluate the result by using it: the form either submits or it does not, the list either sorts or it does not.
It works badly for anything whose correctness is not visible from the outside — pricing logic, permissions, anything where a plausible wrong answer looks exactly like a right one. There, someone has to read what was produced.
Related terms
- Prompt-to-app
- A tool that takes a description of an application and returns a running application, rather than a code sample you then have to host yourself.
- Agentic coding
- A model that plans, edits and re-runs code across several steps instead of answering a single question. Vibe coding is what it feels like to use one; agentic coding is how it works underneath.
- Full-stack generation
- Generating the backend, data model and frontend together, so the result runs, rather than generating a UI that has nothing behind it.
What holds up and what does not
Good for the first version
Getting from an idea to something people can click is where the approach is strongest. You find out whether the idea is worth building before spending a week on it.
Good for software nobody would fund
The internal tool one team needs, the tracker for one process. Real work that never clears the bar for a developer's time.
Weak without a way to check the result
A model will happily produce something that looks right. If you cannot tell whether the output is correct, describing it more confidently does not help.
Questions about vibe coding
No. No-code tools give you a fixed set of building blocks and a visual editor to arrange them; you are limited to what the blocks do. Vibe coding produces actual code, so the ceiling is different — but so is the failure mode, because there is no visual editor showing you what you built.
Not to get started, and that is what makes the term interesting. It helps enormously when something goes wrong, because knowing roughly how software is put together is what lets you describe a problem in a way the model can act on.
It varies, and the honest answer is that "good" depends on what happens next. Code that runs a tool for one team for a year is good enough. The same code underneath a product with paying customers is a liability, and at that point someone has to read it.
Describe an app and see what happens
Caffeine apps run on the Internet Computer, a public network built for this.