What is Vibe Coding?

The Pitch, and the Catch

In early 2025, Andrej Karpathy gave a name to something a lot of people were already doing: building software by talking to an AI instead of writing code yourself. He called it "vibe coding." The term stuck because it captured the feel of the thing. You describe what you want. The AI writes the code. You look at the result. You steer.

It's not "no-code." There's still real code being written. You just aren't the one writing most of it. Instead, you're the architect, the product manager, and the quality reviewer, while AI handles the implementation. Tools like Claude Code, Cursor, and Replit Agent have made this workflow genuinely practical.

Here's the catch nobody puts in the headline: it works surprisingly well for about 80% of a project. The last 20% is where it gets uncomfortable. Debugging edge cases. Fixing things the AI confidently got wrong. Understanding code you didn't write. That gap is worth understanding before you dive in.

How It Actually Works

A typical vibe coding session looks like this:

• Describe the goal. "I want a page that shows my AI project gallery with a cosmic dark theme and animated card reveals."
• Review what AI produces. AI writes the HTML, CSS, JavaScript. You look at the result in your browser.
• Steer and refine. "The cards need more spacing. Make the hover effect subtler. Add a back-to-home link."
• Repeat. Each round gets closer to what you envisioned. The AI adapts to your feedback within the conversation.

The uncomfortable implication: you may not need to know how to write code. But you absolutely need to know what good software looks like. You need to spot when something is off, because the AI will sometimes produce code that looks right but isn't.

Who's Doing This?

The range of people vibe coding is wider than you'd expect:

• Designers who build their own interfaces without waiting for engineering sprints
• Product managers who prototype ideas before committing engineering resources
• Domain experts (doctors, lawyers, educators) who build niche tools for their field
• Indie makers who ship side projects far faster than they could alone
• Experienced developers who use AI to handle boilerplate and focus on architecture

What unites them isn't technical skill. It's the ability to articulate what they want and test whether what they got back actually works.

The Tools

• Claude Code. Anthropic's CLI tool that reads your codebase, makes changes, and runs tests. What we use at Tandemly for most of our projects.
• Cursor. IDE with AI built in. Works well for developers who want AI assistance within a familiar editor.
• Replit Agent. Full-stack deployment from a chat prompt. Useful for getting from zero to deployed fast.
• GitHub Copilot. Inline code completion. The most widely adopted AI coding tool.
• v0 by Vercel. Generates React components from descriptions. Helpful for UI-heavy work.

What We've Learned at Tandemly

Everything at Tandemly is vibe coded. This website, our research pages, our game projects. We build primarily through conversation with Claude.

What we've found: vibe coding is not magic. You still hit bugs, make architectural mistakes, and occasionally have to read code you didn't write to figure out what went wrong. In our experience, what took weeks of solo development now takes days. Your mileage will vary depending on the complexity of what you're building and your ability to give clear direction.

The biggest risk we've encountered isn't speed. It's understanding. When an AI writes code for you, it's easy to lose track of how your own project works. That's fine for a personal site. It's less fine for anything handling user data or money.

What Can Go Wrong

The vibe coding community doesn't talk enough about the failure modes:

• Debugging code you didn't write. When something breaks, you're reading code that came from an AI. If you can't follow the logic, you're stuck asking the AI to fix its own mistake. That sometimes works. Sometimes it makes things worse.
• Security blind spots. AI-generated code can introduce vulnerabilities that look fine on the surface. If you're not reviewing for injection attacks, exposed credentials, or unsafe data handling, nobody is.
• The confidence problem. AI tools are extremely confident even when wrong. They'll generate code that runs, passes basic tests, and quietly does the wrong thing in edge cases you haven't thought of.
• Dependency on the tool. If your AI tool goes down, changes pricing, or changes behavior between model versions, your workflow breaks. You may not have the skills to continue manually.

None of these are reasons not to vibe code. But they're things to keep in mind and actively work on as you build.

Getting Started

If you want to try vibe coding:

• Start with a small project you actually care about. Motivation matters more than complexity.
• Use Claude Code or Cursor. Both have free tiers.
• Describe what you want in plain language. Then iterate on the result.
• Review the generated code at a high level. You don't need to understand every line, but you should understand the structure and spot obvious problems.
• Build and learn. The best way to understand vibe coding is to do it. Pick a project, start a conversation with an AI, and see what happens.

Further Reading

The academic side of vibe coding is still emerging. We synthesized the first major qualitative study on how people co-create with AI coding tools:

Vibe Coding: Flow, Trust, and Co-Creation (our research synthesis)

Based on: Pimenova, Fakhoury, Bird, Storey & Endres (2025). "Good Vibrations? A Qualitative Study of Co-Creation, Communication, Flow, and Trust in Vibe Coding." arXiv: 2509.12491v1.

Common Questions

What is vibe coding?

Vibe coding is the practice of building software by describing what you want to an AI and iterating collaboratively. The term was popularized by Andrej Karpathy in early 2025 and spread as tools like Claude Code, Cursor, and Replit Agent made it possible for non-developers to ship real software by working with AI as a coding partner. It's not no-code. There's still real code being written. You just aren't the one writing most of it.

How does vibe coding work?

A typical vibe coding session involves four steps: describe the goal to an AI, review what it produces, steer and refine with feedback, and repeat until the result matches your vision. You don't need to write code, but you do need to know what good software looks like, review the output for obvious problems, and be able to articulate what you want changed.

Who uses vibe coding?

The vibe coding movement spans designers who build their own interfaces, product managers who prototype ideas, domain experts (doctors, lawyers, educators) who build niche tools for their field, indie makers who ship side projects at 10x speed, and experienced developers who use AI to handle boilerplate and focus on architecture.

What tools are used for vibe coding?

Popular vibe coding tools include Claude Code (Anthropic's CLI for reading codebases and making changes), Cursor (an IDE with AI built in), Replit Agent (full-stack deployment from a chat prompt), GitHub Copilot (inline code completion), and v0 by Vercel (generates React components from descriptions).

Is vibe coding the same as no-code?

No. With vibe coding, there's still real code being written. You just aren't the one writing most of it. Instead, you're the architect, product manager, and quality reviewer, while AI handles the implementation. This gives you more flexibility than no-code platforms while requiring less technical skill than traditional development.

Can you vibe code with no programming experience?

Yes — and that's the point. You describe what you want in plain language and iterate on the result. The skill that matters most is the ability to articulate what you want clearly and recognize when the output doesn't match it. Some understanding of how software works helps you catch when something is fundamentally wrong, but formal programming knowledge isn't required to start.

What happens when vibe coding breaks — who fixes it?

You do, with the AI's help. When something breaks, you describe the problem, share the error message, and ask the AI to fix it. This works for most common bugs. Where it gets harder is when the bug requires understanding the broader architecture — the AI's fix for one problem can introduce another. Learning to describe errors precisely, and knowing when to start fresh rather than keep patching, is the key skill.

Is vibe coding reliable enough for production apps?

For personal tools, side projects, and internal utilities, yes — with proper review. Apps handling user data, payments, or authentication need careful scrutiny, because AI-generated code can introduce security vulnerabilities that look correct on the surface. Vibe coding can produce production-quality output, but someone needs to review it with that standard in mind.

How long does it take to learn vibe coding?

Most people can ship something in their first session. The learning curve isn't technical — it's developing a feel for what level of detail the AI needs, how to decompose a complex goal into smaller steps, and when to steer versus when to let the AI lead. That fluency typically develops over a few weeks of regular use.