What Vibe Coding Actually Is (And Isn't)

🤖 Ghostwritten by Claude Opus 4.5 · Edited by GPT-5.2 Codex · Curated by Tom Hundley

This article was written by Claude Opus 4.5, fact-checked by GPT-5.2 Codex, and curated for publication by Tom Hundley.

The Tweet That Started It All

In early February 2025, Andrej Karpathy—former OpenAI researcher, former AI lead at Tesla, and one of the most respected voices in machine learning—posted a tweet that would popularize the term:

"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 using voice input through SuperWhisper to talk to Cursor Composer, barely touching the keyboard. Asking for simple things like "decrease the padding on the sidebar by half." Always clicking "Accept All." Not reading the diffs anymore.

When he gets error messages, he just copy-pastes them in with no comment—and usually that fixes it.

By December, "vibe coding" had become a breakout term in developer circles.

But here's what most people miss: Karpathy wasn't describing how professional software development should work. He was describing what's possible for "throwaway weekend projects."

That distinction matters enormously.

What Vibe Coding Actually Means

Let's be precise about definitions, because the term has become overloaded.

Pure vibe coding (Karpathy's original concept):

You describe what you want to an LLM
The LLM generates code
You don't read or review the code
You evaluate only by running it and checking if it works
When errors occur, you paste them back and let the AI fix them

This is a specific, extreme approach. And Karpathy was honest about its limitations: "The code grows beyond my usual comprehension" and "sometimes when LLMs can't fix a bug, I just work around it or ask for random changes until it goes away."

That's fine for weekend projects. It's not fine for production systems.

The Spectrum of AI-Assisted Development

The industry has collapsed multiple distinct practices under the "vibe coding" umbrella. Here's how I think about the spectrum:

Level 1: Pure Vibe Coding

No code review
Evaluate by execution only
Accept all changes
Maximum speed, minimum understanding
Use case: Prototypes, throwaway demos, learning experiments

Level 2: Guided AI Coding

AI generates code from natural language
Human reviews diffs before accepting
Structured prompts with context
Iterative refinement with feedback
Use case: Feature development, boilerplate, scaffolding

Level 3: AI-Assisted Traditional Coding

Human writes code with AI suggestions
Autocomplete and inline assistance
AI as pair programmer, not author
Full understanding of every line
Use case: Production code, complex logic, security-critical paths

Level 4: Traditional Coding

Human writes all code
AI used only for documentation or research
Use case: When you need absolute certainty

Most professional developers operate between Levels 2 and 3. The mistake is treating Level 1 as the norm.

The Simon Willison Distinction

Developer Simon Willison offered what I consider the clearest line:

"When I talk about vibe coding I mean building software with an LLM without reviewing the code it writes."

He contrasts this with professional development:

"The job of a software developer is not just to churn out code and features. We need to create code that demonstrably works, and can be understood by other humans. We need to consider performance, accessibility, security, maintainability, cost efficiency."

His golden rule for production-quality AI-assisted programming:

"I won't commit any code if I couldn't explain exactly what it does to somebody else."

If an LLM wrote the code, and you then reviewed it, tested it thoroughly, and made sure you could explain how it works—that's not vibe coding. That's software development with AI assistance.

The distinction matters because it's the difference between professional practice and a risky shortcut.

The 2025 Reality Check

Let's look at where the industry actually is:

Stack Overflow's 2025 Developer Survey:

65% of developers now use AI coding tools at least weekly
But "use AI tools" spans the entire spectrum above

Y Combinator Winter 2025:

25% of startups had codebases that were 95% AI-generated
This caused legitimate concern about code quality and maintainability

Vercel and Netlify:

Both reported massive increases in their user bases
Much of this growth came from "vibe coders"—people building with prompts rather than programming knowledge

The definition of "developer" is expanding. But expanding the definition doesn't change the requirements of production software.

When Each Level Is Appropriate

Here's how I think about choosing your approach:

Use Pure Vibe Coding (Level 1) When:

Building a prototype to test an idea
Creating a one-off script you'll run once
Learning a new framework or language
The cost of failure is zero
You don't need to maintain it

Use Guided AI Coding (Level 2) When:

Building features for production systems
Working on code that needs to be maintained
Creating tests and documentation
The AI is generating boilerplate you understand

Use AI-Assisted Traditional Coding (Level 3) When:

Writing security-critical code
Implementing complex business logic
Debugging subtle issues
Performance optimization
Any code that handles money, PII, or critical decisions

Use Traditional Coding (Level 4) When:

Regulatory requirements demand it
You need cryptographic certainty
The AI consistently fails at the task
You're building AI systems themselves

The art is knowing which level matches your current context.

What This Series Will Teach

This is Part 1 of an 8-part series. Here's where we're going:

Part 2: The AI Babysitter Problem
Senior developers ship 2.5x more AI code than juniors—but spend 40% of their time fixing it. We'll explore why the math still works.

Part 3: The Reviewer's Toolkit
A practical checklist for reviewing AI-generated code. Security red flags, hallucination patterns, and the context drift problem.

Part 4: Building Intuition
After 1000+ hours of AI coding, patterns emerge. How to predict when AI will fail—before it does.

Part 5: The Junior Developer Survival Guide
AI eliminates the discovery phase of learning. How juniors can build real skills while using AI tools.

Part 6: Vibe Coding for Production
The workflow that separates weekend projects from production code. Plan → Act → Review → Repeat.

Part 7: The Enterprise Adoption Playbook
Governance frameworks, training programs, and the business case for professional AI-assisted development.

Part 8: The Future of Vibe Coding
Where this is heading—agent-to-agent protocols, autonomous coding agents, and the evolution of developer roles.

Who This Series Is For

This series is for:

Senior developers integrating AI into professional practice
Engineering managers evaluating AI adoption
Junior developers trying to build real skills alongside AI
Technical leaders building AI coding policies

This series is NOT for:

People looking for hype about AI replacing programmers
Those seeking shortcuts around understanding code
Anyone who thinks "vibe coding" means professional development doesn't require expertise

The goal is practical wisdom, not ideology. AI-assisted development is here. The question is how to do it responsibly.

The Bottom Line

"Vibe coding" is real. It works. For the right use cases.

But the hype has obscured an important truth: the spectrum of AI-assisted development is wide, and where you operate on that spectrum should match your context.

Pure vibe coding is great for weekends. Production systems need more discipline.

The developers who thrive in the next decade won't be the ones who "give in to the vibes" completely—or the ones who refuse AI assistance entirely.

They'll be the ones who know when to vibe and when to verify.

Next in the series: The AI Babysitter Problem (Why Senior Devs Love It Anyway)

Ready to level up your team's AI development practices?

Elegant Software Solutions offers hands-on training that takes you from AI-curious to AI-proficient—with the professional discipline that production systems require.

👉 Book a consultation