Vibe Coding: If AI Writes the Code, What Does the Developer Do?

What is vibe coding?

Vibe coding is a way of building software not by writing code, but by telling an AI what you want in plain language. You handle the "what", and the AI handles the "how".

The term was popularized in early 2025 by AI researcher Andrej Karpathy. As he described it, when vibe coding you barely look at the code: you tell the AI what you want, check the result, paste back any error message — and keep going. The term spread so fast that Collins Dictionary named it the Word of the Year for 2025.

How does it work?

The process usually looks like this:

Describe — "I need a to-do list where I can add, delete, and mark tasks as done."
AI builds — project structure, code, design.
Test — run it, click around.
Ask for fixes — "The button doesn't work", "Make it bluer", "It breaks on mobile."
Repeat — until you're happy with the result.

An example of a good prompt:

Build a simple expense tracker with Next.js and Tailwind.
- Add expenses: amount, category, date
- Monthly chart by category
- Store data in localStorage
- Minimal design with dark mode
Show me a plan first, then write the code.

Which tools?

Agentic tools — Claude Code, Cursor, Windsurf: work inside your project, read and write files, run commands.
In-browser app builders — Lovable, Bolt, v0, Replit: a working site or app without setting up a dev environment.
Chat assistants — Claude, ChatGPT: for small scripts and individual pieces.

The upside

Speed — from idea to working prototype in hours, sometimes minutes.
Low barrier to entry — non-developers can build their own tools.
Less grunt work — AI writes boilerplate, forms, and simple CRUD fast.
Cheap experiments — testing an idea takes an evening, not a week.

The risks

Not understanding the code — when something breaks, you don't know why, and you end up going in circles with the AI.
Security — exposed API keys, unvalidated input, open database access. AI writes code that works, but not always code that's safe.
Technical debt — as the project grows, the code gets tangled and every change introduces new bugs.
False confidence — a working demo is not a working production app.

Vibe coding vs. AI-assisted programming

These are worth keeping apart. If you read, understand, and review the code the AI writes, that's not vibe coding — it's just programming with AI help. The whole point of vibe coding is not looking at the code.

For a personal project or a prototype, vibe coding is great. For anything with customer data, payments, or lots of users, you need to understand the code.

Tips for doing it well

Take small steps — ask for one feature at a time, not the whole app in one prompt.
Use Git — commit every working state so you can roll back when the AI breaks something.
Ask for a plan first — one line, "show me the plan before writing code", prevents a lot of trouble.
Give context — which framework, what style, what you don't want — be specific.
Check your secrets — make sure .env files, keys, and passwords never end up in the code or on GitHub.
Keep learning the fundamentals — however good the AI gets, you need to understand programming to steer it well.

Wrapping up

Vibe coding doesn't kill programming — it changes the developer's role: from the person typing code to the person responsible for the idea, the architecture, and the quality. Use it freely for prototypes and personal tools, and on serious projects, only accept each line the AI writes once you understand it.