Vibe Coding: The Trend That 78% of Companies Have Already Adopted

In February 2025, Andrej Karpathy -- co-founder of OpenAI, former director of AI at Tesla and one of the most influential minds in artificial intelligence -- published a post that changed the vocabulary of an entire industry. He described a new way of programming: you don't write code, youdescribe what you wantand AI generates everything. He called it "vibe coding".

A year later, the concept left Twitter and entered corporate budgets. Data from McKinsey and Upwork shows that78% of organizations have already integrated AI into their development flows. Vibe coding is no longer an early adopter's curiosity and has become a production tool. And with that came benefits, risks and a heated debate about the future of programming.

This article covers everything: where the term came from, how it works in practice, the real adoption data, the risks proven by research, the tools available and, most importantly, how you can use vibe coding effectively.safe and productivein 2026.

1. What is vibe coding (and who coined the term)

Vibe coding is the practice of creating software by describing what you want in natural language, letting AI generate the code. You don't need to know the syntax of a programming language. You don't need to know frameworks. You describe the desired result and the AI ​​writes the code to get there.

The term was coined byAndrej Karpathyin February 2025. In the original post, he described his experience like this: "There is a new type of programming that I call vibe coding, where you completely surrender to the vibe, embrace exponentials and forget that the code even exists."

Karpathy's statement was not just a casual observation. Coming from someone with his resume -- PhD from Stanford, co-founder of OpenAI, responsible for Tesla's Autopilot -- the phrase carried weight. She gave a name to something that thousands of people were already doing, but had no way of describing.

The difference between vibe coding and traditional programming

Aspect	Traditional Programming	Vibe Coding
Input	Manually written code	Description in natural language
Necessary knowledge	Syntax, frameworks, algorithms	Know what you want to build
Prototyping speed	Hours to days	Minutes to hours
Control over the code	Total	Partial (depends on review)
Learning curve	Months to years	Days to weeks
Risk of vulnerabilities	Commensurate with experience	Major without adequate review

It's important to understand that vibe coding is not "not programming". And program at a higher level of abstraction. Instead of telling the computerasdo something step by step, you saywhatyou want and the AI ​​translates your intention into code. The concept is similar to the evolution of Assembly to high-level languages ​​-- each leap in abstraction has allowed more people to build more things, faster.

2. How it works in practice

In everyday life, vibe coding works like this: you open a tool like Claude Code, Cursor or Windsurf, describe what you want to build and the AI ​​generates the code. You review, adjust the description if necessary and iterate until you reach the result.

Real example: creating an API

In less than 2 minutes, you have a functional API with database, validation and documentation. Without typing a line of code. This is vibe coding in essence.

The vibe coding cycle

1. Describe:you explain what you want, with as much context and specificity as possible
2. To generate:the AI ​​produces the code based on your description
3. Review:you analyze the result, test and identify adjustments
4. Iterate:you refine the description and the AI ​​adjusts the code
5. Validate:testing, security review, confirmation that it works as expected

Step 3 -- review -- is where the difference lies between amateur vibe coding and professional vibe coding. Anyone can ask AI to generate code. Few know how to evaluate whether the generated code is safe, performant and maintainable. And exactly wherespecialized skillsmake a difference.

3. The numbers: 78% of companies have already adopted

Vibe coding is not niche hype. Data from 2025-2026 shows massive and accelerated adoption of AI in software development.

According to the report of theMcKinsey "The State of AI in 2025"combined with data fromUpwork Research Institute, 78% of organizations have already integrated generative AI into at least one development function. This represents a 72% jump from the previous year -- adoption is accelerating, not slowing down.

Additional data

• GitHub:more than 1.8 billion code suggestions accepted via GitHub Copilot by the end of 2025
• Stack Overflow Survey 2025:76% of developers already use or plan to use AI tools in development
• Gartner:prediction that 80% of developers will use AI assistants by 2027
• JetBrains Developer Survey:67% of developers who use AI assistants report significant productivity increases

The most revealing data comes from the users themselves: developers who use vibe coding tools report completing tasks55% fasteron average. Not in trivial tasks -- in real projects, with real complexity, in corporate environments.

4. The 40% salary premium for AI-savvy professionals

Here is the data that should motivate any professional to pay attention: according to theUpwork Research Institute, workers who master AI tools receive asalary premium of up to 40%compared to professionals of a similar level who do not use AI.

This number didn't come out of nowhere. It reflects a simple economic logic: professionals who produce more in less time generate more value for their companies. And the market pays for this value.

What “AI-savvy” means in practice

It is not enough to have an account onChatGPT. AI-savvy professionals are those who:

• Integrate AI into the daily flow:They don't use "sometimes", they use it systematically for each type of task
• They know how to give precise instructions:understand how to describe what they want so that AI generates useful results the first time
• Critically review:do not accept AI output blindly -- validate, test and refine
• Choose the right tools:they know when to use Claude Code, when to use Cursor, when not to use AI at all
• Use advanced skills and configurations:costmize their AI tools with specialized instructions that elevate output quality

The salary premium does not go to those who use AI as a toy. It goes to whoever uses it asprofessional multiplier. And the difference between the two groups is in the depth of use -- which is exactly what quality skills provide.

5. The real risks: vulnerabilities and technical debt

Not everything in vibe coding is productivity and acceleration. Recent research reveals concrete risks that every professional needs to be aware of before adopting the practice.

1.7x more serious issues

A study published in 2025 analyzed code repositories generated predominantly by AI and found that projects with high use of vibe coding present1.7 times more issues classified as "major"compared to traditionally developed projects. Major issues are those that affect critical functionality, cause data loss or break main user flows.

2.74x more security vulnerabilities

The most alarming data comes from security research: AI-generated code without adequate review contains2.74 times more vulnerabilitiesof security. This includes SQL injection, cross-site scripting (XSS), exposure of sensitive data and authentication failures.

The illusion of competence

Perhaps the most subtle risk is psychological. Vibe coding makes anyoneopinionable to program. The code works, the project runs, everything seems correct. But beneath the surface there may be problems that only manifest themselves at scale, under load, or when an attacker finds the vulnerability.

This phenomenon -- which researchers call the "illusion of competence" -- is dangerous because the professional doesn't know what he doesn't know. He looks at the working code and assumes that everything is ok. Months later, he discovered that the database had no indexes, that the API did not validate inputs or that user data was exposed.

Accelerated technical division

AI-generated code tends to solve the immediate problem without considering the project architecture as a whole. Each piece of code works individually, but as the project grows, the lack of architectural coherence creates technical debt that accumulates exponentially.

In traditional projects, technical debt accumulates over months or years. With irresponsible vibe coding, it can accumulate in weeks. Speed, which is the main advantage, becomes the vector of the problem.

6. The debate: Simon Willison and the divided community

Vibe coding divided the technology community into well-defined camps. One of the most articulate andSimon Willison, creator of Datasette and historic contributor to Django.

Willison is not anti-AI -- he uses AI tools extensively and documents their use publicly. His critique is specific: he argues that the term "vibe coding" normalizes the practice ofnot understanding the code you are putting into production. For Willison, there is a fundamental difference between using AI as an accelerator (where you understand and review everything) and "vibes" (where you accept the output without understanding).

Both sides

In favor of vibe coding:

• Democratizes software creation -- more people can build solutions
• Dramatically speeds up prototyping and validation of ideas
• Lowers the barrier to entry for non-technical entrepreneurs
• Enables domain experts (marketing, finance, healthcare) to create specific tools without relying on devs
• The market has already voted: 78% adoption does not lie

Against vibe coding (or concerns):

• Code without understanding and code without maintenance
• Security vulnerabilities are a concrete risk
• The "illusion of competence" can lead to dangerous decisions
• Accelerated technical debt costs more in the long run
• Excessive dependence on AI can atrophy fundamental skills

The truth, as in most debates, lies in the middle. Vibe coding is a powerful tool that, like any powerful tool, can be used for good or bad. The question is notseyou should use, butasyou must use.

7. How to do vibe coding securely

The answer to the risks of vibe coding is not to stop using it. And use withprocess, review and appropriate tools. Here is a practical framework for secure vibe coding:

1. Use specialized skills

Skills transform a generalist AI into a specialist AI. When you use a web security skill, for example, Claude Code already generates code with input validation, data sanitization and protection against common attacks. You don't need to remember to ask -- the skill ensures that best practices are in the code from the beginning.

2. Never accept output without review

Read the generated code. Even if you don't understand 100%, look for suspicious patterns: user data exposed in logs, SQL queries assembled by concatenating strings, endpoints without authentication, secrets hardcoded in the code.

3. Automated Tests

Ask the AI ​​to generate tests along with the code. Tests do not guarantee that the code is perfect, but they guarantee that it does what it is supposed to do. And when you change something, the tests break and let you know.

4. Use CLAUDE.md for permanent context

The fileCLAUDE.mdat the root of his project he gives permanent instructions to Claude Code. Include security rules, code standards, and constraints for your project. Every time you order something, Claude Code will respect these rules.

5. Human code review for production

For projects that go into production -- especially if they involve user data, payments, or sensitive information -- always have a second person review the code. Vibe coding speeds up creation, but does not eliminate the need for revision.

8. Tools for vibe coding in 2026

The ecosystem of tools for vibe coding has matured significantly. Here are the main options in 2026, with their strengths and limitations:

Claude Code

terminal toolAnthropic. Largest context window on the market (1 million tokens), direct access to the file system, command execution and extensibility via skills. It is the most powerful option for complex projects and for those who want total control. It works with minhaskills.io skills to increase the quality of output in specific areas.

Cursor

Code editor based onVS Codewith integrated AI. Friendly visual interface, support for multiple models (Claude, GPT,Gemini) and features like "Composer" to generate entire projects. Excellent for those who prefer graphical interfaces and already use VS Code. Supports.cursorrulesfor design instructions.

Windsurfing

Another AI-first editor, focused on "flows" -- sequences of automated actions. Good integration with design and prototyping tools. Positioned as an intermediate option between the simplicity of Cursor and the power of Claude Code.

GitHub Copilot

The pioneer of code wizards. Integrates directly with VS Code, JetBrains and Neovim. Excellent for autocomplete and inline suggestions as you type. Less suitable for "pure coding vibe" (generating entire projects from scratch) and more for traditional development acceleration.

Tool	Best for	Maximum context	Extensible
Claude Code	Complex, terminal projects	1M tokens	Skills (.md)
Cursor	Visual interface, VS Code users	Varies by model	.cursorrules
Windsurfing	Automated flows	Varies by model	Flows
Copilot	Autocomplete, inline	Limited	VS Code Extensions

The choice depends on your profile. If you are a technician and want maximum power, Claude Code. If you prefer graphical interface, Cursor. If you want to autocomplete as you type, Copilot. In practice, many professionals combine two or more tools.

9. When to use and when not to use

Vibe coding is not the answer to everything. Knowing when to use it and when not to use it is just as important as knowing how to use it.

Use vibe coding to:

• Prototypes and MVPs:When you want to validate an idea quickly, vibe coding is unbeatable. In hours you have something functional to test with users
• Built-in tools:automation scripts, internal dashboards, productivity tools. The risk is low and the speed pays off
• Landing pages and websites:Static or semi-static pages are ideal for vibe coding. The generated HTML/CSS is easy to review and the security risk is minimal
• Data analysis:scripts to process spreadsheets, generate graphs, cross-reference data. You validate the result by looking at the numbers, not the code
• Learning:Generating code to study is one of the best ways to learn. You ask, analyze what was generated and understand the "how"
• Marketing automations:integrations with APIs, email scripts, report generation. Repetitive tasks that do not justify hiring a dev

Avoid vibe coding (or use with extreme caution) for:

• Payment systems:Any code that handles real money needs rigorous human review, extensive testing and security auditing
• Sensitive health data (LGPD/HIPAA):Regulations require specific standards of protection that AI may not follow correctly
• Critical infrastructure:Systems that cannot fail (aviation, energy, hospitals) are no place for experimentation
• Cryptography and security:Security implementations need experts. AI can generate something that “seems” safe but has subtle flaws
• Massive scale systems:Performance at scale requires architectural decisions that generalist AI may not optimize correctly

The general rule is:the greater the consequence of an error, the greater the human supervision must be. Vibe coding is an accelerator, not a substitute for professional judgment.

10. The future of AI development

Looking at where the market is heading, some trends are clear for the next 12-24 months:

Skills as a competitive differentiator

As more people adopt vibe coding, the advantage is no longer into useAI, but inasyou use. Specialized skills -- sets of instructions that raise the quality of output in specific areas -- are becoming the differentiator between amateur and professional coding vibes. Those with better skills produce better results, faster. THEskills marketIt has already surpassed 700,000 options in 2026, and is growing exponentially.

Increasingly capable models

With each new generation of models, the generated code gets better. Models like Claude Opus 4 already demonstrate a capacity for architectural reasoning that did not exist a year ago. The tendency is for vibe coding risks (vulnerabilities, technical debt) to decrease as models evolve -- but never completely disappear.

"Assisted" development as standard

By 2027, writing code without any AI assistance will be as rare as writing text without a spell checker. Not because people don't know how to program, but because there is no reason to do manually what AI does in seconds. The debate will shift from “should I use AI?” to “how do I use AI responsibly?”

Regulation and responsibility

With mass adoption comes regulation. The EU AI Act already defines requirements for AI systems in high-risk contexts. It is only a matter of time before specific regulations for AI-generated code emerge, especially in regulated sectors such as finance and healthcare.

The professional of the future

The most valuable professional will not be the one who knows how to program better nor the one who knows how to give better prompts. It will be the one who matchestechnical mastery, critical judgment and strategic use of AI. Someone who knows when vibe coding is the right answer, when it needs manual review, and when AI shouldn't be used.

This professional is not being born from scratch. He's training now, as he experiments with tools, installs skills, tests limits, and develops intuition about when to trust and when to question the AI's output. If you are reading this article, you are already on this path.