What Is Vibe Coding? Matthew Berman Explains His Workflow

SUMMARY

In 30 seconds

Matthew Berman describes vibe coding as telling an AI what to build, trying the result, and iterating without reading the generated code or reviewing diffs. He shares his Windsurf and Superwhisper setup and uses a Python Tetris game as an example.

1A voice-driven workflowBerman describes telling AI what to build and letting it write the code.
2Tools in his stackHe uses Windsurf as his editor and Superwhisper for voice-to-text, and mentions Cursor as an alternative.
3Test and iterateHis example asks for a Python Tetris game. He tests the result and describes problems or pastes errors back instead of inspecting code or diffs.

Takeaways

Berman describes vibe coding as directing AI to write code rather than writing it yourself.
His setup pairs the Windsurf editor with Superwhisper voice transcription; he also mentions Cursor as an alternative editor.
He tests the output and describes problems or pastes errors back to the AI, without reading the code or reviewing diffs.
He says the approach is used by technical people too, citing Andrej Karpathy.

TopicsVibe Coding Coding agents Developer workflows & testing

FULL TEXT

Read the edited text by chapter

Show transcript

Transcript only

Defining vibe coding

Berman describes asking AI to build software without writing the code yourself. He says technical people, including Andrej Karpathy, use the approach.

Vibe coding is a way of coding where you do not write a single line yourself. You use your voice to tell AI exactly what you want built, and it builds it for you. It is not just for nontechnical people: Andrej Karpathy uses it and loves it, and now I do too.

The tools and an example

He names Windsurf and Superwhisper in his workflow and mentions Cursor as an alternative. He illustrates the process with a request for a Python Tetris game.

Here is my stack: I use Windsurf as my code editor, although you can use Cursor, and Superwhisper for voice-to-text. I simply say what I want built—for example, “Build me the Tetris game in Python”—hit enter, and let it go to work.

Testing without reading code

Berman says he tries the resulting software and reports problems or pastes errors into the AI. He describes trusting the AI without reading code or reviewing diffs.

SLIDE 1 / 1Windsurf suggests a dependency commandThe editor shows an AI response about the Tetris project and a proposed command to install its dependencies.

The key is that you do not read the code. You try to use what it produces and see if it works. If it does not, explain what is wrong and what you want built; if there is an error, copy and paste it directly into the AI. You do not look at the diffs: you just vibe and trust the AI to build what you want.