Andrej Karpathy coined the phrase in February 2025, and it stuck faster than anyone expected. The idea is simple enough to explain in a sentence: describe what you want in plain language and let the model write the code.
"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." — Andrej Karpathy
What it looks like in practice
The boring definition: you state the requirement in natural language, an assistant writes the implementation, you look at the result, and you refine by asking for changes rather than editing files.
A real session runs roughly like this.
You: build me a to-do app, dark theme, mobile responsive, saved in local storage.
Assistant: writes a complete HTML, CSS and JavaScript file.
You: good, but let me drag tasks to reorder them.
Assistant: pulls in a sortable library and rewires the list.
You: add categories — work, personal, shopping.
Assistant: adds tags and filter buttons.
Half an hour, a working app, not a line typed by hand. That part is real and it is genuinely useful.
The tools people actually use
- Cursor — a VS Code fork, still the most common starting point
- Claude Code — Anthropic's command-line agent
- GitHub Copilot — autocomplete plus chat, already in most teams
- Windsurf — Codeium's agent-style editor
- v0 — Vercel's UI generator, good for front-end scaffolding
- Bolt — full-stack apps generated from a prompt
Where it works, and where it does not
Being specific matters more than being enthusiastic here.
It works well for prototypes, weekend projects, landing pages and straightforward create-read-update-delete screens. Anything where being wrong is cheap and being fast is valuable.
It is workable with supervision for production features with real business logic — provided somebody reads every line before it merges.
It is a poor fit for security-sensitive systems, performance-critical code, and large refactors of legacy code the model cannot see all of at once. In those situations the confident-but-wrong failure mode is expensive.
The risks nobody mentions in the demo videos
- Confident bugs. The code compiles, reads well and does the wrong thing. That is harder to catch than code that obviously breaks.
- Security gaps. Unescaped input, missing authorisation checks, secrets left in the source. All of it looks fine at a glance.
- Licensing. Generated code can resemble licensed source closely enough to matter for commercial work.
- Skill erosion. If you never debug anything yourself, you gradually lose the ability to — and you only find out during an incident.
- Maintenance debt. The worst outcome is a codebase nobody on the team understands. It works until it does not, and then nobody knows why.
The habit that makes it safe: draft, then verify
The professional workflow that has settled in is not "let the AI do it" or "do it all by hand". It is: let the model produce the draft, then read it line by line as if a stranger wrote it, ask for tests, and deliberately push the edge cases.
That review step is where the value actually lands. Skipping it does not save time — it moves the time to a worse moment, usually in production.
We use these tools in our own work every day. Everything that reaches a client's site is still read by a person first, and that is not a formality.
Does this replace developers?
The boilerplate end of the job — scaffolding, simple screens, repetitive glue code — is already being absorbed. The judgement end — architecture, debugging something strange under pressure, deciding what not to build — is not.
We wrote a longer take on that in what AI actually changes for developers.
If you want to try it this evening
- Install Cursor — the free tier is enough to get the idea.
- Open the chat panel.
- Ask for something small and complete: a Pomodoro timer in React, with dark mode and a sound when the timer ends.
- Then keep refining it, and read what changes each time. That reading is the actual lesson.
Need something built properly for production? Talk to us. We are happy to use AI in the process and equally happy to explain every line we ship.
FAQs
Can vibe coding make me a developer without learning to code?
Not on its own. You still need the fundamentals — otherwise you cannot tell working code from code that merely looks right. What it does do is speed up learning considerably, because you can read a working implementation of anything you can describe.
Is it safe to use AI-generated code in production?
Only with review. The common failure is code that is confident and wrong — it compiles, reads cleanly and does the wrong thing. Read every line, ask for tests, and check the edge cases before it merges.
Which tool should a beginner start with?
Cursor, because it is a VS Code fork and everything you already know still applies. Once you are comfortable, a command-line agent like Claude Code is worth trying for larger multi-file work.
Need a custom solution?
Instacode builds production-grade software — algo trading, ecommerce, web apps. Let's talk.
Get in Touch
💬 Comments (0)
Be the first to comment 🚀