What vibe coding means
In ordinary programming you type the code yourself. In vibe coding you write a request, such as "add a sign-up form that checks the email address", and a language model produces the code. A language model is the kind of AI behind chat assistants: it predicts text, and code is text.
Most tools go one step further and give the model an agent role. An agent is a model that can act: it reads your files, edits them, runs commands and reads the output, in a loop, until the task is done. You steer by reading what it did and replying.
The name comes from the feel of it. You judge the software by using it, and you guide the work by describing what feels wrong, instead of reading every line.
Who it suits
- People with an idea and little coding experience. A founder, designer or analyst can get a working first version without learning a framework first.
- Developers who want speed. Routine work such as forms, tests, scripts and page layouts goes much faster when you review instead of type.
- Anyone exploring. Trying three versions of a feature costs minutes, so you can compare ideas by running them.
It suits less well anything where a hidden mistake is costly: payments, medical or legal logic, security code. You can still use AI there, but someone who understands the code has to read it.
How to do it well
The same few habits come up in every project that goes well.
- Start with a short brief. Write down what the app is for, who uses it, and what the first version must do. Give this to the model before any code is written.
- Work in small steps. One feature per request. A small change is easy to check and easy to undo.
- Run it every time. Code that has not been run is a guess. Open the app, click through the new part, and read any error messages.
- Keep notes outside the chat. A long conversation loses early decisions. A plain file in the project that records what was decided, and how to run things, fixes this.
- Save checkpoints. Use Git, or a tool with built-in undo, so that a bad change costs you a minute and not an afternoon. Git is the standard tool for recording versions of a project.
- Paste errors back in full. The exact error text is the most useful thing you can give a model.
- Ask for an explanation. When you do not understand a change, ask the model to explain it in plain words before you accept it.
Where it fails
| Problem | What it looks like | What helps |
|---|---|---|
| It forgets | The model undoes a decision you made an hour ago | A memory file in the project that every session reads |
| It never checks | The reply says "done" but the page is broken | Run the app after every change, or use a tool whose agents test their own work |
| It sprawls | Features pile up with no structure | A written plan with phases and tasks |
| It hides risk | Passwords in the code, no input checks | Ask for a security review, and have a person read sensitive parts |
| It loops | The same fix is tried again and again | Stop, go back to a checkpoint, and describe the problem differently or try another model |
None of these are reasons to avoid vibe coding. They are reasons to bring structure to it. The more the project grows, the more that structure matters.
Choosing a tool
There are three broad kinds of tool. Browser app builders need no setup and suit a first web app. AI code editors run on your computer and show you the files. Terminal agents work from the command line and suit people who already live there. Our tour of AI coding tools goes through each kind and who it fits.
The model matters as much as the tool. A strong model plans better; a fast one is cheaper for routine edits. See how to choose a model for coding.
How VibeCode.surf approaches it
VibeCode.surf is a desktop AI code editor, in beta, built around the habits above. Agents keep a plan in PLAN.md and a project notebook (AGENTS.md plus a few files in .vibecode/) that every agent reads and that is updated after each working turn. They open your app in a real browser to check their work, and a checkpoint is saved before and after every turn so anything can be undone. The vibe coding guide covers this in more detail.
It is honest to say what it lacks. There is no autocomplete from language servers, no inline AI edits or tab completion, no debugger with breakpoints and no extension marketplace yet. It needs a free account; your projects, chats and model keys stay on your computer. macOS is the tested platform, and the Windows and Linux builds are new. You can install it with one command.