Vibe coding is a software development practice where you describe what you want to build in plain English and accept AI-generated code with minimal manual inspection, relying on follow-up prompts and testing to guide the result toward a finished product. Coined by Andrej Karpathy in February 2025, the term was named Collins Dictionary's Word of the Year for 2025 and has become one of the defining concepts of AI-assisted development. By early 2026, 92% of US developers use AI coding tools daily and roughly 41% of all code written globally is AI-generated.
Where it came from — Karpathy's original post
On February 2, 2025, Andrej Karpathy — co-founder of OpenAI and former Director of AI at Tesla — posted about a new way he'd been building software: '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.' He described accepting AI suggestions wholesale, copy-pasting error messages straight back to the model, and not reading the code the AI produced. His conclusion: it was great for throwaway weekend projects, but 'the code grows beyond my usual comprehension.'
Within weeks it had spread beyond developer Twitter. Merriam-Webster listed it as a trending expression in March 2025. By summer 2025, 25% of Y Combinator startups reported codebases that were 95%+ AI-generated. By year-end, Collins Dictionary named it Word of the Year, reflecting cultural validation that AI-assisted building had become a mainstream expectation, not a curiosity.
Scale of adoption by early 2026
According to surveys across US developers: 92% use AI coding tools daily; 82% of global developers use them at least weekly; 41% of all code written globally in 2025 was AI-generated; 84% of developers report using or planning to use AI tools; and top engineering teams report 85–90% daily AI assistant usage rates.
The vibe coding workflow
| Stage | Traditional coding | Vibe coding |
|---|---|---|
| Ideation | Sketch architecture, choose frameworks, estimate timeline | Write a single high-level prompt describing the entire app |
| Implementation | Write every function, class, and module manually | Accept AI-generated code; describe changes in plain English |
| Debugging | Read stack trace, trace through logic, isolate bug | Paste the error message into the AI; accept its fix |
| Review | Read every line before merging — you wrote it | Test outputs and UX; code internals may be a black box |
| Deployment | Configure servers, CI/CD, environment variables | "Vibe deploying" — one-click deploy buttons in AI-native tools |
A realistic vibe coding session — what the prompts actually look like
== Prompt 1 ==
Build a full-stack web app where users can track their daily habits.
Stack: Next.js, Tailwind, Supabase for auth and database.
Features: add habits, mark them done each day, streak counter,
calendar heatmap view. No need for teams or sharing.
== [AI generates ~400 lines of code] ==
== Prompt 2 ==
The streak counter is showing 0 even when I mark habits done.
Fix it.
== [AI diagnoses and patches the logic] ==
== Prompt 3 ==
The calendar looks ugly on mobile. Make it scroll horizontally
and show month labels. Keep everything else the same.
== [AI refactors the calendar component] ==
== Prompt 4 ==
[paste error] TypeError: Cannot read properties of undefined (reading 'map')
at HabitList.tsx line 47
== [AI fixes the null-check] ==Tools ecosystem in 2026
| Tool | Best for | Skill level needed |
|---|---|---|
| Bolt.new | Fast React/Next.js web app prototypes; one-click Netlify deploy | Zero — no install |
| Lovable | Full-stack apps with real-time multiplayer vibe sessions; backed by a16z | Beginner |
| v0 (Vercel) | UI component generation, Next.js apps; tight Vercel integration | Beginner–Intermediate |
| Replit Agent | Browser-based IDE + agent; free tier; great for students | Beginner |
| Cursor | AI-native IDE for existing codebases; agents that read your whole repo | Intermediate |
| Claude Code | Terminal-first agent; writes, edits, runs, and tests code autonomously | Intermediate–Advanced |
| GitHub Copilot (Workspace) | Enterprise-grade; integrates with existing GitHub workflows | Intermediate |
The quality problem is real
A December 2025 CodeRabbit analysis of 470 open-source PRs found AI co-authored code had 1.7× more major issues than human-written code — including 75% more logic errors and 2.74× more security vulnerabilities. A METR randomized controlled trial found experienced developers were 19% slower with AI tools despite believing they were 20% faster. For production code, especially in healthcare, finance, or security-critical applications: always read and understand the AI-generated code. Vibe coding without review is for prototypes.
How to vibe code responsibly — the spectrum
| Mode | Description | Good for | Risk level |
|---|---|---|---|
| Pure vibe | "Forget the code exists" — accept everything, test the UX | Solo weekend projects; throwaway prototypes; demos | High — never put in production |
| Guided vibe | Review AI output at the function level; test before accepting | Side projects; internal tools; MVPs | Moderate |
| Supervised vibe | AI drafts, human reviews every file, runs tests, checks security | Startup MVPs; production web apps with low stakes | Low-moderate |
| AI-augmented engineering | AI assists specific tasks; human architect drives all decisions | Enterprise; healthcare; finance; regulated industries | Low — professional standard |
The key skill vibe coding teaches
Vibe coding does not eliminate the need to understand software — it shifts the bottleneck. The new critical skill is prompt architecture: structuring your intent clearly enough that the AI builds what you actually want. Developers who understand how the generated code works will always produce better outcomes than those who purely accept and click. "Vibe coding with comprehension" is the professional sweet spot.
Practice questions
- What is the difference between vibe coding and traditional AI-assisted coding (GitHub Copilot)? (Answer: GitHub Copilot: autocomplete assistance — the developer writes code, Copilot suggests completions. Developer maintains full understanding and control. Vibe coding: the developer describes what they want in natural language, the AI generates entire features or applications. Developer may not read the generated code carefully. The distinction is in intent and oversight: Copilot augments developers; vibe coding replaces some development decisions with AI generation. Critics argue vibe coding creates maintenance and security risks from unreviewed AI code.)
- What are the security risks of vibe coding when building production applications? (Answer: (1) SQL injection: AI-generated database queries may lack parameterisation. (2) Authentication bypasses: AI may implement insecure session management. (3) Dependency vulnerabilities: AI suggests popular packages that may have known CVEs. (4) Hardcoded credentials: AI may include example credentials that get committed. (5) Missing input validation: AI generates 'happy path' code without edge case handling. The developer who did not read the generated code cannot catch these issues. Security auditing is mandatory before deploying vibe-coded applications.)
- What context should you provide to an AI coding assistant for best vibe coding results? (Answer: (1) Technology stack and versions ('Next.js 14, TypeScript, Prisma, PostgreSQL'). (2) Existing code structure (paste relevant files or describe architecture). (3) Constraints ('no external API calls', 'must work offline'). (4) Quality requirements ('include error handling', 'add TypeScript types', 'add JSDoc comments'). (5) Security requirements ('use parameterised queries', 'sanitise user inputs'). Without context, AI generates generic code that requires significant adaptation.)
- What is the 'boomerang context' problem in extended vibe coding sessions? (Answer: As a coding session extends, the conversation context grows large and becomes internally inconsistent. Earlier decisions conflict with later ones. The AI may generate code inconsistent with what was generated 20 messages ago. The developer who does not read carefully may not notice. Solution: periodically commit and summarise, start fresh conversations for new features, explicitly provide a project summary to new conversations, and use tools like Cursor that maintain a codebase index rather than relying solely on conversation context.)
- What skills must a developer retain when using vibe coding extensively? (Answer: (1) Architecture reasoning: ability to evaluate whether the AI's structural decisions are sound. (2) Security review: ability to spot common vulnerabilities in generated code. (3) Debugging: when AI-generated code fails, the developer must diagnose root causes. (4) Performance analysis: AI may generate correct but inefficient code. (5) Testing: writing tests that verify AI code meets requirements. Vibe coding does not eliminate the need for software engineering judgment — it shifts the developer from implementation to direction, review, and quality assurance.)
On LumiChats
LumiChats is itself a vibe-coding-friendly environment — you can build entire apps by describing them in the Agent Mode, which uses a sandboxed WebContainer to generate, run, and fix code iteratively without needing to leave the chat.
Try it free