Natural Language Programming (Vibecoding)
Natural Language Programming: what works, what doesn't, and where humans are still needed.
In 2025, Andrej Karpathy, AI researcher and co-founder of OpenAI, coined a term that shook the developer world: vibecoding. Instead of typing code line by line, you describe your intent in plain language and let AI do the implementation. A REST API that used to take two hours? Ten minutes of prompting plus thirty minutes of review.
But this new power comes with a paradox and three dangerous traps. This article explains what vibecoding is, how the tools work, and why the developers who use it best are the ones who could code without it.
The Paradigm Shift
Definition
What Is Vibecoding?
Analogy
Definition
The Next Abstraction Step
- 1. Machine Code
- 2. Assembler
- 3. C / Java
- 4. Python
- 5. Natural Language
Your role shifts: from code writer to code reviewer. The development cycle becomes an iterative dialogue: Prompt → Code → Test → Feedback → Repeat. Development as conversation, not dictation.
Roughly speaking: AI handles about 80% of implementation (boilerplate, standard patterns). You handle the critical 20%: architecture decisions, edge cases, security, and review. The real value lives in that 20%.
2 hours: writing routes, controllers, models, validation, and error handling by hand.
40 minutes: 10 min prompt iteration for a working draft, then 30 min critical review and architecture adjustment.
Caution
Misconception: Vibecoding Makes Programmers Obsolete
Vibecoding shifts the core task from writing to reviewing, architecture, and requirements engineering. Without programming knowledge, you can't evaluate whether AI code is correct, secure, and maintainable.
The Tool Spectrum
AI coding tools operate at different autonomy levels — from line-level autocomplete to fully autonomous agents. The choice depends on the task, project size, and desired control. Most tools now cover several levels, up to an agent; below I sort them by what they became known for.
Four Levels of Autonomy
GitHub Copilot: The AI pair programmer
The democratization of AI-assisted software development for millions of developers. On June 29, 2021, GitHub announced the technical preview of Copilot – an AI pair programmer, powered by OpenAI Codex. Based on a GPT-3 variant trained with billions of lines of public code from GitHub repositories, Copilot could generate code completions and entire functions from comments. The underlying Codex model achieved a 28.8% success rate on first attempt in the HumanEval benchmark – significantly better than GPT-3's 0%. Particularly impressive: With 100 sampling attempts, the success rate increased to 70.2%. Copilot worked especially well with Python, JavaScript, TypeScript, Ruby, and Go. The limited technical preview generated enormous interest and established AI-assisted programming as a viable tool. Copilot fundamentally changed the developer experience and paved the way for a new generation of AI-powered coding tools.
Caution
Misconception: The Most Autonomous Tool Is Always Best
For security-critical code, complex state management, or performance-sensitive paths, lower autonomy with tighter human control is often safer. Match tool autonomy to task risk.
Vibecoding has a surprising consequence: you write less code but must read more code — and read it more critically than ever. Understanding AI-generated code is harder than understanding your own, because you lack the mental model of how it was created.
Code reading becomes the core skill. Accepting AI code blindly builds technical debt that becomes expensive later. The paradox: vibecoding doesn't make programming skills obsolete — it makes them more important.
Interactive: Build a Vibecoding Prompt
Assemble a typical vibecoding prompt from building blocks. Select a role, expertise (tech stack), task, and output format (quality requirements) — and watch how a structured prompt emerges that gives your AI tool clear guidance.
Build a system prompt from building blocks. Click on a block to add or remove it.
Your System Prompt
Good system prompts follow a pattern: define a role, specify expertise, describe the task, and set the output format. By assembling from building blocks, this pattern becomes visible and the effect of each part becomes tangible.
Risks and Limits
Vibecoding carries three critical risks that demand constant vigilance. On top of that, a security dimension: AI happily implements SQL injection vulnerabilities, missing input validation, and exposed secrets.
Caution
Three Critical Risks
1. Hallucinated Code: Compiles correctly but contains subtle bugs. Edge cases are ignored, false assumptions and security holes baked in. Sometimes the AI even invents packages that don't exist, and attackers can claim exactly those names for malicious code. Blindly accepted AI code is about as risky as code blindly copied from Stack Overflow.
2. Context Loss: On large projects, AI loses track. It forgets internal conventions, misses hidden dependencies, and repeats errors that were previously fixed.
3. Cargo-Cult Programming: Accepting code you don't understand creates "vibecoding debt" — a codebase nobody can maintain. The longer review is skipped, the higher the invisible debt.
A beginner generates a chatbot app. The local demo works perfectly. In production: hardcoded API keys in frontend code (security breach), no rate limiting (cost explosion), no error handling (crashes on network issues).
Everything compiled. Everything "worked" — until real-world conditions exposed the missing fundamentals. A study by the research institute METR (early 2025) shows how easily you overestimate yourself here: experienced developers were 19% slower with AI tools, yet believed they were 20% faster. METR itself now calls the result outdated: a follow-up study with newer tools (February 2026) points more towards a time saving, but gives no reliable figure.
Caution
Misconception: If the Code Runs and Tests Pass, It's Safe
AI-generated code often compiles and passes basic tests but ignores edge cases, security requirements, and architectural best practices. Testing and review are non-negotiable.
The bottom line: the best vibecoders are experienced developers who know what excellent code looks like. Vibecoding amplifies skill — it doesn't replace it.
Key Takeaways
Note
Key Takeaways
- Vibecoding is intent-based programming: you describe WHAT, AI delivers HOW. It's the next abstraction layer after high-level languages.
- Your role shifts from code writer to code reviewer and architect. The roughly 80% that AI handles frees you for the 20% that creates real value.
- AI coding tools span a spectrum from autocomplete to autonomous agents. Match the tool's autonomy to your task and desired control level.
- Three risks demand constant vigilance: hallucinated code, context loss, and cargo-cult programming. The paradox: vibecoding makes you write less code but requires you to read more code.
Check
Quiz: Vibecoding
What does the term "vibecoding," coined by Andrej Karpathy, describe?
Select one answer
1. What does the term "vibecoding," coined by Andrej Karpathy, describe?
2. A developer needs to add authentication to a large enterprise application. Which tool autonomy level is most appropriate?
3. An AI coding tool generates a login form that compiles, renders correctly, and passes basic tests. A senior developer discovers passwords are stored in plain text. Which vibecoding risk does this illustrate?
4. A junior developer argues: "Since AI writes all my code now, I don't need to learn to read code anymore." What is the fundamental flaw?
5. Which tasks fall into the critical 20% that humans must handle in vibecoding?
6. A startup builds its entire MVP using vibecoding. Six months later, no team member can explain how the authentication module works. What is this situation called?
Check
Comprehension Check
Can you explain this in your own words? Tick what you've got.
0 of 3 done