What Is “Vibe Coding”?

Vibe coding is a new, informal term for writing software by talking to AI tools instead of typing code. In practice, a person simply describes a task in plain English (or another natural language) to a large language model (LLM) that’s been trained to write code. The AI then generates working source code to solve the task. As one summary explains, it shifts “the programmer’s role from manual coding to guiding, testing, and refining the AI-generated source code”.

In other words, you tell the computer what you want, and it (mostly) writes the program. AI researcher Andrej Karpathy, who introduced the term in early 2025, described it as “not really coding, I just see stuff, say stuff, run stuff, and copy-paste stuff, and it mostly works” . The name comes from Karpathy’s viral social-media post. He said vibe coding means you “fully give in to the vibes” and even “forget that the code even exists” .

In practical terms, this can look like talking or typing simple instructions (e.g. “make the button green,” “filter out these items,” “suggest lunch options based on my fridge photo”), and the AI updates the code behind the scenes. Some people even use voice commands (Karpathy demonstrated speaking to an AI tool via a transcription service) so they “barely even touch the keyboard” . The key idea is that anyone can start building software just by conversing with an AI assistant, without needing to learn the syntax of Python, JavaScript, or any traditional language.

Example: In one common scenario, a user might prompt ChatGPT (or a similar model) with, say, “Write Python code to sort a list of numbers.” The AI responds with a code snippet. The user then runs it, sees an error or a needed change, and simply tells the AI what to fix. This loop of prompt–generate–test–refine is at the heart of vibe coding .

Vibe coding is already entering the dictionary. In March 2025 Merriam-Webster listed it as a “slang & trending” noun. Of course, it’s slang, not a rigorously defined term, but it captures a real trend: everyday users experimenting with AI as a co-programmer. Vibe coding is often compared to no-code or low-code development, but with AI doing the actual code-writing. In effect, English (or another spoken language) becomes the user’s programming language. As Karpathy quipped earlier in 2023: “the hottest new programming language is English”

How AI Tools Enable Vibe Coding

At its core, vibe coding depends on modern AI “coding assistants.” These range from general chatbots to specialized IDE plugins and cloud tools. Popular examples include:

ChatGPT and Claude: Large AI chatbots (like OpenAI’s ChatGPT or Anthropic’s Claude) are trained on vast amounts of code. You can prompt them in plain language, and they will output whole functions or programs. For instance, ChatGPT can generate Python scripts, JavaScript apps, or SQL queries just from your description. Users often “chat” with the model, asking follow-up questions or requesting tweaks, as if conversing with a junior developer.
GitHub Copilot (and Agent Mode): This is an AI coding extension inside your code editor (like Visual Studio Code). Copilot suggests lines of code as you type, and in newer “Copilot Agent” modes it can respond to natural-language commands (e.g. “Create a login page” or “Optimize this function”). It effectively auto-completes your thoughts in code. As one early vibe coder noted, using these tools has multiplied his personal coding efficiency by about five times.
Interactive AI Coding Platforms: New platforms like Cursor AI, Replit’s Agent, Windsurf, Vibe.Dev, Lovable.dev, and others have been built specifically for this style. They typically let you describe an entire application or feature in English and then generate the full front-end and back-end code. Some can even import a design (from Figma) and turn it into a working web app automatically. For example, Cursor’s “Composer” feature can chat with you about code changes and apply them directly, while Replit offers an “AI Agent” that you can instruct in plain text.
Voice and Multimodal Tools: As Karpathy demonstrated, voice recognition can be combined with AI coding agents. Using tools like SuperWhisper, he literally spoke to the AI assistant, saying things like “decrease the padding on the sidebar by half,” and the AI edited the code for him . In principle, any interface (voice, typed prompt, or even image inputs) could be used to “vibe code.”

These AI systems are essentially supercharged autocomplete on steroids. A typical workflow might be:

Describe the feature: Say (or type) what you want the program to do. For example: “Create a webpage that displays my to-do list.”
AI writes code: The LLM generates code (HTML/JS, Python, etc.) implementing that feature.
Run and review: You run the code. If something is off or you get an error, you again describe the problem to the AI (for example, “Make the font bigger” or copy-paste the error message and ask “What went wrong here?”).
AI fixes or iterates: The model updates the code accordingly.
Repeat until done: This loop continues until the tool does what you want. At each step, the user isn’t typing code by hand – they are guiding the AI by saying changes in plain language.

These steps create a conversational, iterative “AI-mediated” development process, rather than the traditional coding loop of writing and debugging code manually. The image below illustrates someone using ChatGPT to generate code in real time. A user interacting with an AI coding assistant (shown here is ChatGPT 4o) that generates Python code from natural-language prompts. This is an example of vibe coding in action .

Vibe Coding in Action: Examples

Vibe coding has already shown up in some real projects and experiments. Tech writers and hobbyists have tried it out:

“LunchBox Buddy” by Kevin Roose (NYT): In early 2025, technology columnist Kevin Roose, who is not a trained programmer, used vibe coding to build several small apps. One was LunchBox Buddy, which analyzes photos of the contents of his fridge and suggests items for a packed lunch. Roose describes these AI-generated tools as “software for one” – highly personalized apps that solve a specific personal need. This example even made it onto Wikipedia’s front page as a Did You Know? fact: “a journalist used vibe coding to create an app to suggest what to pack for lunch”. The embedded screenshot above comes from that Wikipedia page.
Engineering Case Studies (IEEE Spectrum): An article in IEEE Spectrum interviewed three engineers experimenting with vibe coding. None were fresh computer-science grads: two were an experienced data engineer and a mechanical engineer. Yet with AI they built software he never could have before. For example, mechanical engineer Prasad Naik converted an old C program (for an iPad) into a modern web app in about two hours. He said he learned JavaScript on the fly by asking ChatGPT how to do each part. In the end “more than 90 percent of the code he used was generated by ChatGPT,” and he admitted “I can’t explain in detail how the AI-generated code…functions. But it works”.

Similarly, Jason Touleyrou, a data engineer, used AI to quickly build a cloud-based coffee-tracking app using Google BigQuery and Pub/Sub – services he had never used before. His point: AI let him prototype without wasting time. “Speed to ideation is critical,” he said. In effect, he had “ideas I wanted to test” and AI turned them into working code in minutes or hours, whereas they would have taken days or weeks by hand.

Startup Development (Y Combinator): Vibe coding is also hitting the startup scene. Y Combinator (a top tech incubator) reported in March 2025 that about 25% of its Winter batch startups had codebases that were 95% AI-generated. YC’s president Garry Tan has been a vocal proponent. He says vibe coding lets very small teams achieve what once needed dozens of engineers. In his words: it allows a startup of “10 people” to do the work “that might take a team of 50–100 software developers”. He claims such AI-assisted companies have “getting to a million… or 10 million dollars a year revenue with under 10 people” – something that “really never happened before” . Tan advises struggling young engineers to “vibe code and build startups” and even suggests that by mastering tools like Cursor or Windsurf, a small “fully vibe coder” team could “literally do the work of 10 or 100 engineers in the course of a single day” .

These stories show the enthusiasm around vibe coding. Enthusiasts emphasize how it dramatically lowers the barrier to creating software. Suddenly, beginners or domain experts (marketing folks, analysts, hobbyists) can build custom apps without learning syntax. As one LinkedIn user put it, vibe coding is popular “because for the first time…people with little or no coding experience can create…sophisticated coding programs and projects just using vibe coding, just speaking to AI in natural language”. They point to real benefits:

Speed and Productivity: With AI handling routine coding, a person can iterate on ideas faster. One coder noted that using AI tools “increased my coding efficiency by about five times”. Startups are launching MVPs (minimum viable products) in days instead of months.
Learning Acceleration: By asking the AI for code, newcomers see examples of working code on the fly. It’s like learning by example instantly. IEEE Spectrum even suggests vibe coding can be an “accelerator for gaining new skills” – a front-end engineer could quickly become “full-stack” by prompting AI to generate server code and then learning from it.
Creativity and Fun: Many report that the process feels playful. You’re essentially “programming by persuasion,” and it can feel like a creative conversation. Andy Gordon (a data advisor) described feeling awe at the AI’s “sorcery” while also the satisfaction of guiding it to a solution through clever prompting.
Democratization: Experts often say this is a democratizing force. Just as smartphones put powerful tools in millions of hands, vibe coding puts software creation in reach of non-engineers. You could be a teacher, a chef, or a hobbyist with an idea, and the AI can help you build the tool to realize it.

Cautions and Criticisms

Despite the buzz, many caution that vibe coding has serious limitations. The excitement tends to focus on quick prototypes or “toy” projects. For anything mission-critical, reliance on AI-generated code can be risky:

Lack of Understanding: By definition, vibe coding means you often accept AI-written code without fully understanding it. This can hide bugs or security issues. Wikipedia notes that developers “may use AI-generated code without fully comprehending its functionality, leading to undetected bugs, errors, or security vulnerabilities”. In Roose’s experiments, for example, one AI-generated program “fabricated fake reviews” for an e-commerce site – an obvious mistake that a careful coder would not have introduced. These kinds of hallucinations are a known flaw of LLMs.
Unpredictable Quality: AI models often produce working code, but it might not follow best practices. It may be verbose, inefficient, or brittle. One Reddit commenter noted that AI could “overcomplicate some formulas” or produce code that looks correct but fails under different conditions. Without solid knowledge, the user might not know how to improve or optimize this code.
Maintaining Real Code: Vibe coding works when building something from scratch, but professional software is usually an ongoing project. Simon Willison (a veteran developer) warned: “Vibe coding your way to a production codebase is clearly a terrible idea.” He points out that real engineering work is maintaining and extending existing code. In that context, code readability and understanding are vital. If nobody knows what the code is doing (it’s just “vibes”), then fixing bugs or adding features later becomes very difficult. Willison argues that if you do review and understand each AI-generated line, then it’s not really vibe coding anymore – it’s just using AI as a smart autocomplete tool.
False Sense of Expertise: Because vibe coding can let non-experts make something, there’s a danger of thinking they know what they made. As one critic quipped, vibe coding is like hiring a “drunk uncle” to build a race car kit and then taking credit for the build. Beginners might call themselves “programmers” after a few AI-assisted projects, even though they haven’t learned core principles of coding or debugging. This can backfire: when something goes wrong (and it inevitably does), the lack of foundational knowledge means the user may not be able to fix it.
Difficulty of Prompting: Surprisingly, some find that getting the AI to do exactly what you want can be quite challenging. Taher Vohra, a 25-year veteran engineer, says he tried the strict vibe-coding approach (never even looking at the code) and found that “specifying what I want the AI to do is turning out to be a harder problem than doing it myself.”. In other words, writing the right prompts can be as tricky as writing the code. And if the prompt is unclear, the AI’s output may go astray.
Long-Term Risks: Experts warn about pushing prototypes into production too quickly. A small project that seems fine with AI-generated code might become a nightmare if it’s expanded. IEEE Spectrum notes the risk that “a good enough prototype often faces pressure to get pushed to production.” If the underlying code was written by AI and not fully vetted, it could contain hidden technical debt. Checking and testing every corner of AI code takes effort, and without that, subtle bugs can slip in.

These criticisms don’t mean vibe coding will vanish. Instead, they suggest it should be used carefully. Most proponents agree it’s great for hobby projects, learning, and proof-of-concept work. But they caution it should not replace skilled engineers when building secure, mission-critical systems. As Simon Willison put it, “for experiments and low-stake projects… go wild! But stay aware of the very real risk that a good enough prototype often faces pressure to get pushed to production”.

The Tools of Vibe Coding (AI in the Loop)

The rise of vibe coding has been fueled by a surge in AI-powered developer tools. Here are some key categories:

Chat Interfaces: Models like OpenAI’s GPT-4 (ChatGPT), Anthropic’s Claude (Sonnet), and Google’s Gemini can be accessed via chat. You simply describe what you need and the model replies with code. Users often keep a running conversation, refining the program step by step. (Tip: Save your conversation, since each reply might depend on the previous prompts.)
IDE Assistants: Plugins like GitHub Copilot (and its new “Agent Mode”) or Amazon CodeWhisperer suggest code snippets as you type. They can also respond to natural language prompts: “Write a function to reverse a string,” for example. This lets you stay in a development environment while still vibing with the AI.
Dedicated Web Apps: Tools like Cursor AI, Vibe.Dev, Replit AI Agent, Windsurf, Lovable.dev, and Base44 provide visual interfaces. Often you start with a blank app and instruct the AI: e.g., click a button in the UI to “Add Feature”, then describe the feature. The platform generates both the front end and back end code, which you can preview instantly.
Voice/Multi-modal Tools: Integrations like Whisper or SuperWhisper transcribe voice prompts into text for the coding assistant. This is still experimental, but it’s been demonstrated. Imagine saying “Make the logo bigger” to your AI partner and immediately seeing the change. In the example, Karpathy connected Whisper to Cursor so he could program “without even touching the keyboard” .

Many of these tools can be run for free (at least on small projects) or offer free trials, so beginners can experiment with vibe coding without cost. As one developer noted, using a mid-tier coding AI is like having a “Lamborghini body with a Fiat 500 engine” if the model isn’t very good. The performance depends on the AI model’s capabilities. (Right now, models specialized for coding – e.g. OpenAI’s Codex or Anthropic’s Sonnet – tend to give better results than general-purpose ones.) Popular AI Coding Tools (examples):

OpenAI ChatGPT or Anthropic Claude (accessible via browser chat or API)
GitHub Copilot / Copilot Agent (IDE extension)
Cursor AI (chat + coding agent)
Replit AI (browser-based coding + deployment)
Amazon CodeWhisperer (IDE/VSCode plugin)
Tabnine (autocomplete with AI)
Custom setups: connecting GPT-4o or Claude to speech, or even to images (e.g. generating code from a drawn sketch)

The rapid tooling evolution is itself a part of the vibe coding story. One Medium blogger describes a “mini-industry” of vibe-coding tools, and notes that many of them existed before Karpathy’s tweet, but the hype gave them all a catchy category name. The general trend is clear: more and more development platforms are offering some form of conversational AI interface. Even enterprise IDEs are exploring “AI pair programmer” modes.

Pros and Cons of Vibe Coding

Breaking it down, here are some commonly cited advantages and concerns:

Pros:

Accessibility: Non-programmers can create useful software by describing it in their own words. This might empower small businesses or individuals to build tools without hiring a developer.
Speed: Quickly prototyping ideas. Vibe coding can shave development time from weeks to hours for a simple app.
Learning Aid: Beginners can learn programming concepts by seeing AI-generated code in context.
Lean Teams: As Garry Tan noted, small teams can “stay leaner” by using AI to do bulk coding tasks .
Experimentation: It’s fun and encourages experimentation. You can try ideas with little commitment, since even “throwaway weekend projects” are easy to spin up.

Cons:

Hidden Bugs: AI code can fail silently or do unintended things. Without understanding it, you may not notice.
Lack of Clarity: The resulting code might be hard to read or maintain, since it was never really designed by a human.
Overreliance: Users might become too dependent on AI and never learn the fundamentals. (As one critic put it, vibe coding “only makes sense if you have very current developers looking at it”.)
Security and Quality Risks: Experts warn vibe coding is risky for production systems. Simon Willison says developing via vibe coding “defies conventional wisdom” that code must be high-quality and understandable. Another developer asked skeptically: if AI doesn’t understand your code, how can it diagnose subtle issues?
Prompt Complexity: Figuring out how to phrase prompts can be an art. A vague prompt yields wrong code; the user may need to experiment just to get the AI on track. This “prompt engineering” becomes part of the skill set.

In summary, vibe coding is great for prototyping and personal projects – building that one-off utility or testing an idea. But it is cautioned against use for critical systems without review. As Ars Technica reports, one developer found his AI assistant even refused to write production code for him (it responded, “I cannot generate code for you, as that would be completing your work, you should develop the logic yourself”). This may reflect ethical guardrails in tools, or simply that AI “knows” its limits.

Vibe Coding Today: A Buzzword and a Reality

By 2025, “vibe coding” became a Silicon Valley buzzword. Business press wrote about it constantly. Business Insider declared it “Silicon Valley’s next act,” a way to let “you fully give in to the vibes”. They quoted tech CEOs and investors forecasting big change: Y Combinator’s Garry Tan says AI lets startups “stay leaner” and build fast. OpenAI’s Sam Altman predicted software engineering would look “very different by the end of 2025” (reflecting this trend). Meta’s Mark Zuckerberg even said on Joe Rogan’s podcast that AI would soon be doing the work of mid-level engineers.

Meanwhile, enthusiasts have started calling themselves “vibe coders” on social media, posting demos of games or apps they built by prompting AI. Influencers on YouTube and LinkedIn share step-by-step videos of them modifying code purely by issuing voice commands or text prompts. For the curious non-programmer, it feels a bit magical: “I had little to no experience in coding…in the last two weeks [I was] able to create a website and also a food-planning app” using vibe coding, wrote one user.

But there’s also backlash. On forums and Q&A sites, skeptics criticize the term itself as “cringe” and say vibe coding “makes no sense” if the user doesn’t deeply understand what’s happening. Some accuse others of slapping together AI code without pride or skill. Voices in these communities have a point: as one engineer remarked, using AI without knowledge is like “knowing how to use a hammer but not how to actually build a boat”. Critics stress that if you can’t read the code, you can’t learn from it or ensure it’s correct.

Those debates aside, a few practical trends are emerging:

Learning and Training: Many see vibe coding as a tool for learning new technologies. If a web developer has never used, say, Rust or Swift, they can ask an AI to generate a small program and then inspect it. The IEEE article quotes a veteran saying vibe coding is “an accelerator for gaining new skills”, e.g. moving from front-end to full-stack in months instead of years.
New Roles: The role of a “programmer” is subtly shifting. You still need thinkers who can break down problems and test solutions. But some of the grunt coding work can be offloaded. Some talk about developers evolving into “product engineers” or “architects” who oversee AI-written code. The Slack and GitHub discussions find that vibe coding is essentially another layer of abstraction.
Non-coders Code: In business and research, non-technical people are already using AI to do small coding tasks. For example, data analysts use Copilot to write spreadsheet formulas or SQL queries, or marketers generate email templates with AI. The model’s role is like an expert assistant who does the typing. As one user said, vibe coding is “like having a ChatGPT by your side to quickly generate code stubs and answer questions”.
Industry Adoption: Big companies are experimenting too. Y Combinator revealed that ~81% of its current startups are AI-centric, and many of those rely on LLMs to build their products. Businesses like Replit and startups like Cursor have raised funding specifically because this vibe coding idea caught on. In other words, what started as a meme is now influencing where VC dollars flow.

What Does the Future Hold?

Vibe coding is still very new, and it’s uncertain how it will evolve. But there are some educated guesses:

Democratization vs. Expertise: For non-coders, vibe coding could be revolutionary. Imagine a teacher creating a classroom app, or a small-business owner automating a task, all by talking to AI. That’s empowering. It echoes the no-code movement but potentially with more flexibility. However, experts argue that a basic level of coding literacy will remain important. You still need to understand concepts like data flow and logic to spot errors. If not, you risk blindly trusting the AI’s output.
Changing Skillsets: The emphasis may shift from writing syntax to designing prompts and tests. Good vibe coders will learn to articulate needs precisely, to review AI output critically, and to combine AI-written pieces. Over time, we might teach programming more as “how to work with AI” instead of “how to write loops.” (Some educators are already experimenting with this approach.)
Job Impact: Big tech leaders believe many routine coding jobs will change. If 10 engineers with AI can replace 50-100, then entry-level roles might shrink while demand grows for those who can oversee AI systems or work on very complex problems. Some even suggest mid-level coders might become obsolete, though senior architects and devs who understand core systems will still be needed. It’s analogous to how calculators affected accounting: everyday calculation became easy, but accountants moved up to analysis.
Quality and Governance: Companies will develop governance around AI-generated code. We might see tools that automatically check AI code for bugs or security holes. Development teams will likely combine vibe coding with traditional best practices: for instance, using linters, code reviews, and tests even on AI-crafted code. Some already warn against “Accept All” blindly – that’s a recipe for disaster if used carelessly.
Legal and Ethical: As AI scrapes existing code to train, there are copyright issues. If a vibe-coded program accidentally reproduces copyrighted logic, who is responsible? This is an open legal question. Also, if an AI assistant injects biased or malicious code, liability could become complex.

Possible Scenarios

To ground these ideas, here are a few hypothetical (but realistic) scenarios of vibe coding in the future:

Personal App Builder: Alice wants a budget tracker. She logs into a web-based AI coding assistant. She types, “Create an app to track income and expenses, with charts.” The AI spins up a basic web app. Alice then asks, “Add a button to input transactions by scanning receipts.” The AI adds that feature. In an hour, Alice has a working prototype that would have taken her days to code by hand.
AI-as-Developer Assistance: Bob is a junior developer working on a team. He writes code but struggles with a tricky library. He types an English prompt to Copilot within VSCode, and the assistant generates the function he needs. Bob reads it and understands the logic enough to integrate it. Later, he uses the same tools to fix a security warning. He’s still “coding,” but with heavy AI support.
Citizen Developer at Work: Carla, a marketing manager, has no formal coding background. Her company gives her access to an internal AI agent. Carla tells it to “Generate a web form for collecting survey responses and save data to our company database.” The AI generates the HTML form, JavaScript validation, and backend code to store results. Carla can tweak labels or styles by just telling the AI “Change the header to say ‘Customer Feedback Survey’.” She creates the tool without writing a line of code herself.
Senior Engineer and AI: Dave, a senior engineer, doesn’t need to type every line either. He uses voice commands during a whiteboarding session, saying, “Set up a new REST API endpoint for orders.” The AI agent writes the boilerplate. He then says, “Now write the query to find the five most recent orders,” and the AI completes it. Dave reviews the code, makes one small adjustment, and moves on to architecture. His focus is on high-level design, trusting AI for routine parts.

Each scenario is a different flavor: hobbyist, junior dev with help, non-tech employee “coding,” and expert with AI horsepower. In all cases, the common thread is talking in plain language and getting code.

Conclusion

Vibe coding is a fascinating and somewhat controversial development in programming. It has the potential to make software creation more accessible, faster, and even fun. Already, people with little formal training are experimenting with AI assistants to build custom apps and tools. At the same time, seasoned developers and industry experts emphasize that the vibes alone can’t guarantee correctness. Code quality, security, and understanding still matter.

For the future, it’s likely that we will see a hybrid model: using vibe coding for initial drafts, prototypes, and helper tasks, while retaining traditional development practices for robustness and scalability. Early evidence suggests the most successful approach is partnership: humans define the problem and check the solution, and AI does the typing in between.

As one engineer put it, vibe coding is great for “low-stake projects where you want to explore what’s possible and build fun prototypes”. It might not replace hand-coding entirely, but it’s certainly adding a powerful new tool to every programmer’s toolbox, and making coding itself feel more like a creative conversation. The vibes are here; now it’s up to us to use them wisely.

Sources: