Vibe Coding Course for Beginners

On February 2, 2025, Andrej Karpathy posted a description of a way of working he called vibe coding. The post says: "There's a new kind of coding I call 'vibe coding', where you fully give in to the vibes, embrace exponentials, and forget that the code even exists." He names Cursor Composer with Sonnet. He talks to Composer with SuperWhisper and barely touches the keyboard. He asks for small things, such as "decrease the padding on the sidebar by half," because he is too lazy to find it. He "Accept All" always, and he does not read the diffs anymore. Error messages get pasted back in with no comment, and that usually fixes them. The code grows beyond his usual comprehension, and he would have to really read through it for a while. When the model cannot fix a bug, he works around it or asks for random changes until it goes away. He calls this "not too bad for throwaway weekend projects." Then: "I just see stuff, say stuff, run stuff, and copy paste stuff, and it mostly works."

That post is the definition a beginner course starts from. Collins later named vibe coding its Word of the Year for 2025 and described it as the use of artificial intelligence prompted by natural language to write computer code. Karpathy's wording also says he stops reading the code.

What you practice

Practice that loop on a throwaway project you can look at.

Cursor's quickstart is the first session the product documents now: install, sign in, and open a folder. Open Agent with Cmd+I on a Mac, or Ctrl+I on Windows and Linux. That shortcut is on the Agent mode page. Ask the agent to explain the project before you change anything. The quickstart's prompt is:

Explain this codebase. Point me to the main entry points, key modules, and anything I should read before making changes.

The agent searches the repo, reads relevant files, and summarizes how the project fits together.

Then one small change. The quickstart says good first tasks are low risk, such as copywriting or a small UI fix, and it tells you to ask for options and wait:

Suggest three small, safe improvements in this codebase. Explain the tradeoffs and wait for me to choose one.

If you already know the change, describe the result in one sentence. The padding request is that shape. Run it and look. If an error comes back, paste the message in. That is the repair step in his post. Cursor's reviewing and testing guide says to consider breaking a change that is too large to review comfortably into smaller pieces.

What you still check yourself

His post names the step he left out. He accepts the edits and does not read the diffs. The code grows beyond his usual comprehension. He keeps that to throwaway weekend projects.

The quickstart puts the step in the session. The diff view shows changes as the agent makes them. When it finishes, you review the diff and ask it to run the checks the project already uses: tests, the type checker, linting, or a local build. The Agent mode page says the edits are applied while the agent works, and you reject anything you do not want. Click Stop to halt a run. To undo, hover an earlier message and use Restore Checkpoint. That rolls the files back to that point and leaves the messages in the chat.

The reviewing guide says agent-written code can look right, follow existing patterns, compile, and pass tests, and still miss an edge case, duplicate logic, or carry a security problem. Passing tests can mean the tests are checking the wrong behavior. You still have to understand the change.

Simon Willison, writing about the same post, uses "vibe coding" for building software with a model without reviewing the code it writes. Once you have reviewed the change, tested it thoroughly, and can explain how it works to someone else, he calls the work software development. For anyone just starting, he says keep the project low stakes, watch secrets such as API keys, and watch any service that bills by usage. If other people will use what you built, he says to check with someone more experienced before you share it.

When the change spans many files, press Shift+Tab for Plan Mode. The agent researches the codebase, asks clarifying questions, and writes a plan you approve before it builds. A quick change stays in Agent mode. If the build comes out wrong, revert, make the plan more specific, and run it again.

Where Cursor's docs take over

The post and the quickstart cover the first small change. After that, use the manual.

The Agent mode page defines Ask, Plan, and Debug. Ask is read-only. Plan edits after you approve. Debug is for bugs that need runtime evidence. Switching modes starts a fresh context window, and a new task should start a new chat. Reviewing and testing code covers stopping a run, tagging @Branch to review the whole diff, and asking the agent to write tests you then check. The Plan Mode page covers when a plan is worth writing, and when to revert and edit it.

A course from the post and the quickstart

Ailurn turns a prompt, a PDF, a GitHub repo, or a docs URL into a course you take in the same workspace. It does not watch a YouTube video. Pasting a link does not import a transcript. Start from the AI course builder and ask for a beginner course on vibe coding: Karpathy's definition, then Cursor's quickstart in order (explain the project, one small change, read the diff, and run the project's tests, type checker, linter, or build before you keep the edit). Plan Mode comes after that small change is familiar. Give the builder the quickstart URL as the docs source.