Andrej Karpathy on Vibe Coding: What Developers Should Learn from It

TL;DR

• Andrej Karpathy coined "vibe coding" in a post on X on February 2, 2025.
• In it, he describes talking to an AI coding tool, accepting its changes without reading the diffs, and pasting error messages back, which usually fixed them.
• He called it "not too bad for throwaway weekend projects".
• Our take, not his: the skills this workflow rewards are clear intent, fast iteration and judging the output.
• Also our take: code that other people depend on still needs tests and review.

Introduction

Andrej Karpathy, a founding member of OpenAI and later the Director of AI at Tesla, gave vibe coding its name. He did it in a single post on X on February 2, 2025, describing how he had been building a web app with an AI coding tool.
The post is short and specific, and a lot of what gets said about "Karpathy's approach" goes beyond it. This guide keeps the two apart: first what the post actually says, then what we think developers should take from it. Where a point is our interpretation rather than his, we say so. If you are weighing whether to trust AI-generated architecture at all, our list of questions founders ask before trusting AI architecture is a useful companion.

Understanding Andrej Karpathy Vibe Coding

What the Original Post Says

Karpathy describes vibe coding as a new kind of coding where you "give in to the vibes" and "forget that the code even exists." He says it had become possible because LLMs were getting very good, naming Cursor Composer with Sonnet. He talks to Composer through SuperWhisper, so he barely touches the keyboard, and asks for trivial edits, such as halving the padding on a sidebar, rather than finding the code himself.
He accepts every change without reading the diffs, and when an error appears he pastes it back without comment, which usually fixes it. He sums it up as not really coding: he looks at things, says things, runs things and pastes things, and it mostly works.

The Limits He Pointed Out Himself

The post is candid about the costs. The code grew beyond what he could normally follow without sitting down to read it. Some bugs the model could not fix, so he worked around them or asked for random changes until they went away. And his verdict was a modest one: good enough for throwaway weekend projects.
What the post does not do is set out rules for professional teams or make predictions about the profession. The lessons below are our reading of it, not his instructions. For how we think AI output should be organized once a project gets serious, see our guide on structuring AI-generated code for a real project.

Key Lessons Developers Should Learn from Andrej Karpathy on Vibe Coding

1. Focus on Problem-Solving Instead of Syntax

Our first takeaway: when an AI tool can write the boilerplate, the valuable work moves to deciding what to build and checking that it works. Karpathy's post shows the extreme version, where he barely looks at the code at all.
Most developers will land somewhere in between. In our view, understanding systems, performance and architecture now matters more than memorizing syntax, and so does picking the right vibe coding tools for the job.

2. Learn to Work Effectively With AI Tools

The workflow in the post runs on short loops: ask for a change, run it, paste back the error. That makes prompting, iterating and judging the output the core skills, as we see it, and they can be learned. Be precise about what you want, change one thing at a time, and check the result before asking for the next change.

3. Maintain Engineering Discipline

This lesson is ours, not Karpathy's. He described accepting changes unread, and judged the approach good enough for throwaway weekend projects. For code that customers or teammates depend on, we think testing, code review and a clear structure still decide whether AI-generated code is safe to ship, and that includes checking it for security problems.

How Vibe Coding Reflects the Future of Development

Build this with a prompt on Greta.sh — describe the app, Greta writes and deploys it.→

AI-Native Development Workflows

Integrated AI assistants, automated refactoring and conversational debugging are changing how developers work with codebases. Karpathy's post is an early, informal account of what that feels like from the inside.
Our view: developers who try these tools on small projects first get a better sense of where they help and where they get in the way.

A Shift in Developer Roles

As more code is generated, more of a developer's time goes into design, review and deciding what to build. That is our reading of the trend, not a claim made in the post. The balance to strike is more speed without less reliability.

Practical Steps to Apply Andrej Karpathy Vibe Coding Today

Adopt Iterative Prompt-Based Workflows

The post describes a loop: ask, run, paste the error back, repeat. Start with the same loop on something small and low-stakes, and refine your instructions as you go rather than expecting a perfect first result.

Strengthen Fundamentals While Using AI

Algorithms, system design and performance still matter, because they are how you judge whether generated code is any good. This is our advice; the post does not discuss it.

Experiment With AI Development Tools

Try modern coding assistants, automated testing tools and AI-enabled IDEs on real work. Our walkthrough on building your first web app with Greta is one place to start.

Challenges and Criticism Around Andrej Karpathy on Vibe Coding

Relying on AI output you have not read carries real risks. Developers who stop learning fundamentals may struggle to debug what the AI produces, and code quality needs a review process.
Karpathy named some of these limits himself: the code outgrew what he could easily follow, and some bugs he could only work around. His own verdict was that it was good enough for throwaway weekend projects.

Why Karpathy's Vibe Coding Post Still Matters

The post gave a name to a way of working with AI coding tools, and the name spread fast: Collins Dictionary made "vibe coding" its Word of the Year for 2025.
Its value for developers, in our view, is its honesty about the trade-off: speed and fun on one side, code you do not fully understand on the other. Deciding where a given project should sit on that line is the real skill.

Greta

Greta transforms clear natural language instructions into structured, production-ready applications. Rather than generating isolated snippets, Greta focuses on building systems that follow defined architecture and workflow patterns.
The output includes a complete application stack,frontend, backend, and supporting infrastructure,designed to fit real development environments.
Key capabilities include:

• Prompt-driven application generation with architectural context
• Structured frontend and backend code aligned to project standards
• Built-in workflows that support iteration and refinement
• Exportable source code ready for version control and CI/CD

By emphasizing structure over raw generation, Greta helps teams apply best practices for AI code, maintain consistency, and integrate AI output seamlessly into real projects. This makes Greta especially well-suited for teams looking to move from experimentation to reliable, production-ready development.

Conclusion

Karpathy's post named vibe coding and described it plainly: fast, fun, with code you stop reading, and by his own verdict good enough for throwaway weekend projects. The lessons in this guide (focus on intent and design, get good at iteration, keep tests and review for code that matters) are our interpretation of what developers should take from it. For a closer look at the difference between generating an app and planning a product, see prompt-to-app vs plan-to-product.

FAQs

Want to try one of these? Start on Greta.sh — no setup, no boilerplate — just a prompt.→

1. What does Andrej Karpathy say about vibe coding?

In his February 2, 2025 post on X, he describes letting an AI tool write the code while he accepts its changes without reading the diffs and pastes error messages back. He judged it good enough for throwaway weekend projects.

2. What is Andrej Karpathy AI coding?

It is a search phrase rather than a term from his post. It usually means AI-assisted coding in the style the post describes: you say what you want, and the AI writes and changes the code.

3. Is Andrej Karpathy vibe coding replacing traditional programming?

The post makes no such claim. Our view: it speeds up prototypes and repetitive work, while code that others depend on still needs developers who can read and review it.

4. Why does Karpathy's vibe coding post matter?

It gave a name to a way of working with AI coding tools, and the term spread quickly. Collins Dictionary chose "vibe coding" as its Word of the Year for 2025.

5. How can developers start vibe coding?

Pick a small, low-stakes project, use an AI coding assistant, and work in short loops: ask, run, fix.

6. Does vibe coding require new skills?

In our view, yes: writing clear prompts, judging AI output and thinking about the whole system matter more than before.

7. Are AI tools central to vibe coding?

Yes. The AI writes the code. In the post, Karpathy names Cursor Composer with Sonnet, and SuperWhisper for talking to it.

8. Who benefits most from vibe coding?

Our view: people building prototypes, internal tools and side projects, where speed matters more than long-term maintenance.

9. What is the biggest lesson from Karpathy's post?

Our reading: match the approach to the stakes. Accepting AI changes unread can be fine for a weekend project; code that people rely on needs review.

10. How is vibe coding influencing the future of programming?

Our view: it moves more of a developer's time toward design, review and deciding what to build.