What vibe coding means

On 2 February 2025, Andrej Karpathy, a well-known AI researcher and a co-founder of OpenAI, posted about a new way of working he called vibe coding. You describe what you want, accept whatever the AI writes, paste error messages back when something breaks, and, in his words, "forget that the code even exists". He said it was fine for throwaway weekend projects. The term spread fast, and Collins Dictionary named it its word of the year for 2025.

People now use the term for two different things. The programmer Simon Willison draws the line clearly. Vibe coding is building software with an AI without reviewing the code it writes. Using AI while you still read, test and understand the result is AI-assisted programming. Both are useful. The first is fine for a toy you will throw away. It is risky for anything that holds other people's data or money.

This course teaches a careful version. You can remember it as describe, check, keep control:

• Describe. Write down what you want before you ask, in a form the AI can follow: who uses it, which screens, what data, which rules (A spec the AI can follow).
• Check. After each change, test it yourself the way a user would, and ask the AI to explain anything you don't understand (Reading and fixing what the AI wrote).
• Keep control. Work in small steps you can undo, keep a copy of the code that you own, and decide yourself on anything that touches data, money or security (Building in small steps, Data and security basics).

You do not need to write code to work this way. You need to know what to ask for, what to check, and when to stop and call an engineer.

How the tools work, in plain words

Every tool in this lesson runs on a large language model. A model reads text and predicts the most likely text to come next, one small piece at a time. It learned this from a huge amount of public text and code. When you ask for "a screen with a list of products and a plus button", it writes code that looks like the thousands of similar screens it has seen.

The model only sees what is sent to it with each request. This is called its context: your message (the prompt), earlier messages in the same chat, and the files the tool chooses to show it. Some tools also let it run commands and read what comes back. Three things follow from this, and they explain most AI mistakes:

• It does not know your business. Nobody told it that Ravi buys on 30 days' credit, or that rice is sold only in full bags. It fills the gap with the most common answer it has seen.
• It can be confidently wrong. A prediction comes with no warning label. Code that uses a feature that does not exist looks exactly like code that works, and an explanation of a bug can sound sure and still be a guess. People call this a hallucination.
• It forgets what it was not shown. A new chat starts empty. In a long chat, early details may be summarised or dropped to make room (how this works varies by tool). If a rule matters, write it in a file the tool reads every time, or repeat it.

So the quality of what you get depends on two things you control: what you describe, and what the tool can see. That is why the next lesson is about writing a spec. Developers who want the deeper version, with completion, chat and agents inside a team's workflow, instruction files and reviewing AI pull requests, can read Working with AI coding assistants. Designing AI features explains the models themselves.

Three kinds of tools

Tools change every few months, but in 2026 they fall into three groups. The names below are examples to help you recognise them, not recommendations. Most have a free tier with a daily or monthly limit, and paid plans for one person commonly start at around $10 to $25 a month at the time of writing. Prices, credits and limits change often, so check each tool's current pricing page before you start.

Browser app builders, such as Lovable, Bolt, v0 and Replit. You type what you want into a chat on a website. The builder creates the whole app and shows a live preview next to the chat. Most can also host the app on a public link and connect a database. The builder sees your chat and the whole project, because it made it, and it controls everything, including where the app runs. Your code lives on the builder's servers. To leave, most can connect to GitHub or let you download the code; check how before you build anything you care about. Cost is usually credits per message, so a long afternoon of changes can use up a month's free allowance.

AI code editors, such as Cursor, Windsurf, or VS Code with GitHub Copilot. A code editor is an app on your computer for working with code files, and the AI sits in a side panel. It sees the files you open or point to, and it suggests changes that you accept or reject. Most editors also have an agent mode that makes larger changes and runs commands, usually after asking you. Your code lives in a folder on your laptop, and you can send a copy to GitHub. You install things and run the app yourself, which later lessons show step by step.

Coding agents, such as Claude Code, OpenAI Codex CLI and Gemini CLI. These work in your project folder through the terminal. You give a task in plain words. The agent reads the files it needs, edits several of them, runs commands (install a package, start the app, run checks), reads the result and tries again. It can do a lot in one go, so it needs the most supervision: work on a copy you can throw away, and read its summary of what changed. Many agents now also run inside editors or on the web, so the lines between the three groups keep blurring.

Good at, bad at

The tools are strong where the internet has thousands of examples and where a mistake is easy to spot:

• Screens and forms: a product list, a sign-up form with required fields, a confirmation screen.
• Standard CRUD work, such as adding an item to a usual order, changing its quantity or removing it.
• Wiring well-known services together, such as a common sign-in service or a payment provider's test mode, following that service's public instructions.
• Explaining code in plain words, line by line.
• Fixing errors you paste in, when the error message points at the problem.

They are weak where the answer is not written anywhere they can see, or where a small error costs real money:

• Business rules nobody wrote down. A tool can't know that Ravi's order must stay under his credit limit, or that a shop can only order full cartons, unless you say so.
• Security and access rules. Who may see and change which data is easy to leave out, and the app still looks finished. Data and security basics covers sign-in and access rules.
• Data that must be exactly right, like money and stock. Rounding a price, adding up an order, or taking stock off twice can go wrong in ways a screen never shows.
• Big changes in one go. "Rebuild the order flow with a new design" changes so much at once that you can't check it. Small steps are easier to check and to undo.
• Keeping a big project consistent. The tool sees only part of a large project at a time, so it may add a second way of doing something that already exists.
• Knowing what is current. The model learned from text up to a certain date. It may suggest an old version of a tool or a setting that has since changed.

The pattern behind both lists: an AI tool produces what is likely, and you check what is true. The more an error would cost, the more you check. This is not only theory. In 2025 a security researcher scanned 1,645 apps from one popular builder's public showcase and reported that 170 of them let anyone read personal data from their databases. Their screens worked. Their access rules were missing.

Picking a tool, and the risks you own

Choose by the job in front of you. Most beginners meet three jobs:

Whatever tool you use, four risks stay yours. The tool's company does not own them, and neither does the AI.

• Secrets. An API key works like the key to a building. Never paste one into a chat, and never let it end up in code that runs in the browser, where anyone can read it. Data and security basics shows where keys belong.
• Other people's data. Build with sample data. Real shop names, phone numbers and orders go in only after access rules are checked. Also read what the tool's company keeps from your chats and files, and don't paste a company's private data into a tool the company has not approved.
• Licences. Code and packages the AI adds come with a licence that sets the rules for reuse. Ask the tool to list every package it added and its licence, and check the list before you share the app.
• Cost overruns. Credits, hosting and AI usage can all grow quickly. Set a spending limit where the tool allows one, and look at the bill every week in the first month.