
Vibe coding: building software by prompt, not by reading
In short: Vibe coding means building software by describing what you want to an AI and accepting its code without close review, which suits low-stakes prototypes.
Vibe coding is building software by describing what you want to an AI model and accepting the code it writes, largely without reading or understanding it, steering by whether the result seems to work. Andrej Karpathy coined the term in a post on 2 February 2025 about giving in to the vibes and forgetting the code is even there, a style he called fine for throwaway weekend projects.
Vibe coding in plain words
Ordering food in a language you do not speak by pointing at photos. Great for a quick meal; risky if you have an allergy.
How it works
You prompt an app builder or a coding agent with what you want. It generates code and runs it. You react to what you see ("make the button bigger", "it crashes when I log in") and repeat. You judge by behaviour, and the code itself goes unread.
Example
As an illustration, building a personal habit tracker in an afternoon: describe the screens, paste error messages back when it breaks, deploy when it looks right. For a personal tool, that is a reasonable trade.
Where you'll meet it
- Prototypes, demos and personal tools
- Non-developers building internal apps
- Hackathons and "build an app in a day" challenges
- App builders such as Lovable, Bolt, v0 and Replit
What it changes for you
It lowers the barrier to making software, which is good for prototypes and validation. For professional developers, the line that matters is review: code you ship to others, handle money or data with, or maintain for years needs someone who understands it. For teams, vibe-coded prototypes are useful specs, not production systems.
Vibe coding vs agentic coding
Agentic coding uses AI agents to do much of the work while an engineer still reviews, tests and owns the code. Vibe coding skips the understanding. Simon Willison drew the same line in a March 2025 essay: if you reviewed the code, tested it and could explain it to someone else, it is software development and the label does not apply. The tools can be identical, and the responsibility is what differs.
Common mistake
My view: vibe-coding something with a login and a database, then putting it on the internet. Authentication, data access rules and secrets are where unreviewed code fails most visibly. Willison's essay lists the same hazards: exposed secrets, private data and usage-billed APIs without a spending cap. Prototype by vibe, and bring in review before anything touches real users.
Related terms
Agentic coding · Claude Code · Spec-driven development · Demoability · Plan mode
From the archive: Private app generation still needs an admission decision
Prototype by vibe, and have someone who understands the code review it before it meets real users.


