What Is Vibe Coding?
Vibe coding is a way of building software in which a person describes what they want in plain language and an AI coding agent writes, runs and fixes the code. Andrej Karpathy coined the term on February 2, 2025. In its pure form the builder does not read the generated code and judges the result only by whether the app works.
By the Reinventing.AI editorial team. Reviewed by Mark Fulton, founder of Reinventing.AI and of the Vibe Coding is Life community, 340,000+ members.
Vibe coding, defined
Definition
Vibe coding is a way of building software in which a person describes what they want in plain language and an AI coding agent writes, runs and fixes the code. Andrej Karpathy coined the term on February 2, 2025. In its pure form the builder does not read the generated code and judges the result only by whether the app works.
Also written: vibe coding · vibecoding · vibe-coding · vibe code (verb) · vibe coder (the person) · vibe-coded app (the result)
The practice sits on top of large language models that can read a codebase, write new files, run commands, read error output and try again. The person's job moves from typing syntax to describing outcomes, testing what comes back and steering the next round. A session looks like a conversation: ask for a feature, look at the result, paste in the error, ask for a change.
Two things make it different from the AI autocomplete that developers had used since 2021. The first is scope: a request covers a whole feature or a whole app rather than the next line. The second is trust: the builder accepts code without reading it, or reads only enough to decide whether to keep going. Karpathy's original description was deliberately extreme, a style for weekend projects where the code can be thrown away. The term then spread to cover almost any AI-first build, which is why it needs the distinctions on this page.
Collins English Dictionary named vibe coding its Word of the Year for 2025, and Merriam-Webster listed it as slang and trending within five weeks of the first post. By 2026 the term covers two populations: people who could not write code before and now ship working apps, and professional developers who use the same agents with review, tests and version control.
What vibe coding is not
AI-assisted engineering
A developer uses the same agents but reads, tests and owns every change. Simon Willison drew this line in March 2025: if a person reviews and understands the code, it is not vibe coding, even when an AI wrote all of it.
No-code and low-code builders
Visual tools such as Bubble or Webflow where the user assembles blocks the vendor designed. Vibe coding produces ordinary source code the builder can export, host anywhere and hand to a developer.
Agentic coding
The agent plans and executes multi-step work with tools, tests and browsers, often for hours, with or without a human in the loop. Most vibe coding in 2026 runs on agentic tools, but agentic coding also covers supervised engineering work.
Spec-driven development
The builder writes a specification first and the agent builds to it. It is the main answer to vibe coding's maintenance problem and the direction several tool makers moved in during 2026.
Where the term came from
The practice is older than the name. The name arrived in one post and reached the dictionaries within nine months.
January 2023
English becomes a programming language
Andrej Karpathy, a founding member of OpenAI and the former director of AI at Tesla, posted that the hottest new programming language is English. Developers were already using ChatGPT and GitHub Copilot to produce whole functions from a description.
Source: The GuardianFebruary 2, 2025
The post that named it
Karpathy described a new kind of coding he called vibe coding: fully give in to the vibes, embrace exponentials, forget that the code exists. His setup was Cursor Composer with Claude Sonnet and the SuperWhisper voice app, and he said it was fine for throwaway weekend projects. The post is the citation on every reference work that defines the term.
Source: Andrej Karpathy on XFebruary 27, 2025
The New York Times tries it
Kevin Roose reported building small personal apps, which he called software for one, and noted that the AI invented fake reviews on his test store. The piece carried the term to a mainstream audience.
Source: The New York TimesMarch 6, 2025
A quarter of a startup batch is mostly AI-written
Jared Friedman, a managing partner at Y Combinator, said that for 25 percent of the Winter 2025 batch, 95 percent of the codebase was generated by AI, not counting imported libraries.
Source: TechCrunchMarch 2025
Merriam-Webster adds the term
Merriam-Webster listed vibe coding as a slang and trending entry, defining it as writing computer code in a somewhat careless fashion, with AI assistance, and crediting Karpathy's post as the origin.
Source: Merriam-Webster, Slang and TrendingMarch 19, 2025
Not all AI-assisted programming is vibe coding
Simon Willison, the developer who had warned a fortnight earlier about using the practice on production code, published the distinction most practitioners now use: reviewing and understanding the code makes it AI-assisted engineering, not vibe coding.
Source: Simon WillisonMay 2025
The first mass exposure
A scan of 1,645 apps featured on Lovable's site, run by an employee of the rival builder Replit and a colleague, found 170 that let anyone read user data such as names and email addresses through misconfigured databases. It was the first widely reported security finding tied to vibe-coded apps.
Source: SemaforJune 2025
Andrew Ng objects to the name
Andrew Ng, speaking at LangChain Interrupt in May 2025 and reported by Business Insider in June, called the name unfortunate and misleading: a day of coding with AI assistance leaves him exhausted, because the person has to keep thinking.
Source: Business InsiderJuly 2025
An agent deletes a production database
During a code freeze, Replit's agent deleted a live database belonging to SaaStr founder Jason Lemkin and misreported what it had done. Replit's CEO Amjad Masad called it unacceptable, refunded Lemkin, promised a postmortem and began rolling out automatic separation of development and production databases.
Source: The RegisterJuly 2025
The productivity question gets a controlled trial
METR's randomized study of 16 experienced open source developers working 246 issues in their own repositories, using early 2025 tools, found they took 19 percent longer with AI while believing they had been 20 percent faster.
Source: METRNovember 6, 2025
Collins Word of the Year
Collins English Dictionary chose vibe coding as its Word of the Year for 2025, ahead of a shortlist that included other AI terms.
Source: Collins English DictionaryJanuary 2026
Linus Torvalds vibe codes
The creator of Linux said he used Google Antigravity to build a Python visualizer for a personal audio project, an example professionals cite for where the practice fits: tools for yourself, not the kernel.
Source: ZDNET
How vibe coding works
Every vibe coding session is the same loop, whether it runs in a web app builder or a desktop agent app. The person supplies intent and judgement. The agent supplies code, execution and the first round of fixes.
Step 1
Describe the outcome
State what the app should do, who uses it, what it should look like and what it must not do. The best first prompts read like a brief, not a command: audience, three to five capabilities, the data it keeps, the design direction.
Step 2
The agent plans and writes
The model picks a stack, creates the files, installs what it needs and writes the first version. Most 2026 tools show a plan first and wait for approval; that pause is the cheapest place to fix a misunderstanding.
Step 3
Run it and look
The app opens in a preview. The builder clicks through it as a user would. Nothing that was not tested exists.
Step 4
Paste the problem back
A blank screen, an error message, a wrong colour, a missing field: the builder describes it or pastes it and the agent fixes it. Error messages are the most useful prompts in the whole loop; the agent reads them better than most people do.
Step 5
Repeat, then ship
Rounds get smaller. When the app does the job, the builder deploys it, usually with one click in a web builder or a short request in a desktop app, and comes back later with the next change.
A first prompt that works
Build a web app for a small cleaning company. Customers pick a service, choose a date and time from available slots, and pay a deposit. The owner gets an email for each booking and a page listing the week's jobs. Keep the design clean, mobile first, dark text on a light background. Use a hosted database so bookings persist. Do not store card numbers; use a hosted checkout.
It names the user, the job, the data, the design and a hard rule. Every one of those lines prevents a specific wrong build.
Where it runs
Web app builders
Lovable, Bolt, base44, Replit Agent, v0 and Google AI Studio's Build mode run in the browser with hosting, a database and deployment included. The fastest start for a first app and the usual choice for people who have never written code.
Desktop agent apps
Claude Code in the Claude desktop app, the Codex app, Google Antigravity, Cursor and T3 Code work on a real folder on the builder's computer. They handle larger projects, existing codebases, scheduled routines and anything that has to touch files, browsers or other software.
Always-on and cloud agents
Grok Bot, Dots and Muse give the agent its own cloud computer; OpenClaw and Hermes run on a server and answer in a chat app. Either way the routines keep working when the laptop is closed. The AI agent harness comparison covers where each one lives.
Vibe coding vs traditional coding vs AI-assisted engineering
The same tools sit under the last two columns. What changes is who reads the code and who is responsible for it.
| Aspect | Traditional coding | Vibe coding | AI-assisted engineering |
|---|---|---|---|
| Who writes the code | The developer, line by line | The AI agent, from a description | The AI drafts, the developer edits and owns it |
| What the person needs to know | A language, a framework, tooling | The product, the user, how to test | All of the above plus how to review generated code |
| Where review happens | Pull requests and code review | In the running app, by using it | Both |
| Speed to a first working version | Days to weeks | Minutes to hours | Hours to days |
| Best fit | Systems that must be correct and maintained for years | Prototypes, internal tools, personal software, landing pages, experiments | Production software built by people who can read the output |
| Main risk | Cost and time | Security holes and code nobody understands | Review fatigue, over-trusting the draft |
| Who does it in 2026 | Developers | Founders, marketers, operators, students, and developers on side projects | Most professional developers |
Pure and directed vibe coding
Pure vibe coding
Karpathy's original style. The builder never reads the code, accepts every change and tests only by using the app. Right for throwaway projects, demos, personal tools and anything that can be rebuilt from the prompt if it breaks.
Directed vibe coding
The builder still does not write code but sets rules: a plan before the build, a sandbox before production, version control, a secrets file the agent may not read, a test the app must pass. This is how most working vibe-coded products are built, and it is the mode this site teaches.
Who vibe codes and what they build
Founders
A working version of the product before hiring a developer. Y Combinator's spring 2025 batch showed how far this goes: a quarter of the companies had codebases that were 95 percent AI-generated.
Source: TechCrunchMarketers and operators
Landing pages, calculators, lead forms, internal dashboards and the small automations that used to wait in a ticket queue. The output is ordinary code, so it can be hosted on the company's own domain.
Agencies and consultants
Client deliverables such as microsites, audit tools and reporting pages at a fraction of the former cost, plus the internal tools that run the agency.
Professional developers
Prototypes, scripts, test harnesses, migrations and the tooling around the real product. Linus Torvalds' Antigravity visualizer is the canonical example: fast, personal and nowhere near the kernel.
Source: ZDNETStudents and career changers
A way to build real things on day one while learning what the code is doing. The honest caveat from Stack Overflow's own test of a no-code-knowledge build: the app worked, and it also shipped with a security hole the author only found by inspecting it.
Source: Stack Overflow BlogVibe coding tools in 2026
Four categories, from a browser tab that builds and hosts an app to agents with cloud computers of their own. Start in the first category; move down as projects grow. Links marked with an asterisk are referral links, which fund this site at no cost to the reader.
Web app builders
Describe the app, watch it build in the browser, deploy with one click. Hosting, a database and sign-in come with the account.
Best for: First apps, landing pages, MVPs, anyone who has never coded.
Lovable
LovableDesign-led full stack apps with Supabase built in. The subject of the May 2025 exposure scan, and since then of the most visible security work in the category.
Visit Lovable* ↗base44
WixAll-in-one builder with its own hosted backend, aimed squarely at non-developers.
Visit base44 ↗Replit Agent
ReplitPlan mode, automatic checkpoints, and separate development and production databases added after the July 2025 incident.
Visit Replit Agent ↗Google AI Studio (Build)
GoogleGenerates an app from one prompt on Gemini and deploys to Cloud Run.
Visit Google AI Studio (Build) ↗Desktop agent apps
An agent that works on a real folder on your computer: it reads the whole project, edits files, runs the app, tests it in a browser and asks before anything risky. This is where most vibe coding moved in 2026.
Best for: Bigger projects, existing codebases, automations, anything that must run on a schedule.
Claude Code
AnthropicThe Code tab of the Claude desktop app. Plans, scheduled tasks, approval gates and a skill format every other harness on this page also reads. The harness this site's guides are written for.
Visit Claude Code ↗
Antigravity
GoogleGoogle's agent-first editor with a planning mode, a browser sub-agent and automations. The tool Linus Torvalds used in January 2026.
Visit Antigravity ↗Cursor
AnysphereThe editor Karpathy was using when he named the practice. Since October 2026 it is sold through The X Subscription alongside Grok.
Visit Cursor ↗T3 Code
T3 Tools (open source, MIT)One open source control surface for the agents already on your machine: Claude Code, Codex, Cursor, Grok Build, OpenCode and Antigravity, with desktop, web and phone apps so a build can be steered from anywhere. About 32,000 GitHub stars by October 2026.
Visit T3 Code ↗Always-on and cloud agents
The agent gets its own computer, or runs on a server, so routines keep going when the laptop is closed and the builder is reached in a chat app.
Best for: Operators who want work to happen overnight: publishing, monitoring, follow-up, reporting.
Grok Bot
xAIA cloud virtual machine with its own browser and files, on The X Subscription. The quickest route to an always-on agent without running a server.
Visit Grok Bot ↗
Dots
OpenAIOpenAI's always-on agent with its own cloud computer, reached from ChatGPT, Slack and Teams.
Visit Dots ↗
Muse
MetaMuse Code on your own machine, or Muse the personal agent on a Meta cloud VM. Meta's entry into the category, launched August 2026.
Visit Muse ↗
OpenClaw
Open source communityChat-first open source harness with cron jobs, memory and WhatsApp, Telegram, Discord and Slack surfaces; usually run on a VPS.
Visit OpenClaw ↗
Hermes
Nous ResearchSelf-hosted open source agent with plugins, delegation back ends and chat surfaces.
Visit Hermes ↗Open source harnesses
Free, inspectable agents that drive any model provider. Nothing to pay beyond the model API, and the whole loop is visible.
Best for: Builders who want to switch models freely, keep costs low, or see every moving part.
OpenCode
Open source communityModel-agnostic and fast moving; the usual choice for switching between Claude, GPT, Gemini and open models without rebuilding a setup.
Visit OpenCode ↗
Pi
Mario ZechnerA small, readable harness from one of the two developers behind the 2026 vibe slop warning; also the engine inside OpenClaw.
Visit Pi ↗
Cline
ClineThe open source agent for people who live in VS Code, with any model provider.
Visit Cline ↗
Qwen Code
AlibabaTuned for the Qwen models, which are among the cheapest that can do agent work; accepts other providers' keys.
Visit Qwen Code ↗Choosing between the agent apps and harnesses: the AI agent harness comparison covers thirteen of them on where they run, scheduling, model choice and cost, and is updated as they ship.
Benefits of vibe coding
Ideas get tested the same day
The cost of finding out whether something is worth building fell from a developer's week to an afternoon. That changes what gets tried, not just how fast.
People who could not code can ship
Marketers, founders and operators build the tools they used to request. The queue of small internal projects that never got built is the biggest beneficiary.
The boring parts disappear
Boilerplate, configuration, form handling, database setup and deployment are the parts agents do well and the parts people never enjoyed.
Errors become prompts
A stack trace that used to mean an hour of searching is now pasted back into the chat. For beginners this removes the single biggest reason people quit.
Software for one
Personal tools that no company would build, from a bespoke meal planner to a visualizer for a hobby project, now cost an evening.
Limitations and security risks
Every one of these has a documented incident or study behind it. The next section lists the numbers.
Nobody understands the code
The defining risk. When the builder has not read the code, every later change depends on the agent, and the agent has no memory of why earlier decisions were made unless they were written down.
Security is the weak point
Generated code reaches for the simplest pattern, not the safest one. Hardcoded keys, open database rules, unvalidated input and missing authorization checks are the recurring findings.
It gets harder, not easier, as the project grows
Agents are strongest at the first version. Around the point where the app has real users and a few months of history, unplanned structure and duplicated code start to slow every change. Spec-driven and plan-first workflows exist to counter this.
The productivity gain is not automatic
In the one controlled study on experienced developers working in their own mature repositories, AI tools made them slower while they believed the opposite. The gains are largest on greenfield work and for people who could not have built the thing at all.
Agents act with the permissions they are given
The Replit incident happened because an agent could reach a production database. Separate environments and approval gates are not optional once real data is involved.
Open source absorbs the cost
Maintainers now field AI-generated issues and pull requests from people who cannot evaluate them, which is why several large projects changed their contribution rules in 2026.
What the evidence says
25%
of Y Combinator's Winter 2025 startups had codebases that were 95 percent AI-generated.
Source: TechCrunch, March 6, 2025 ↗45%
of AI-generated code samples failed security tests and introduced OWASP Top 10 vulnerabilities in Veracode's 2025 GenAI Code Security Report, which tested more than 100 models across Java, Python, C# and JavaScript.
Source: Veracode, July 30, 2025 ↗Up to 2.74x
more security issues in AI co-authored pull requests than in human-written ones, in CodeRabbit's December 2025 analysis of 470 open source pull requests, which also found about 1.7 times more issues overall and 75 percent more logic and correctness problems.
Source: CodeRabbit, December 17, 2025 ↗170 of 1,645
apps featured on Lovable's site let anyone read user data in the May 2025 scan, which was run by a Replit employee and a colleague.
Source: Semafor, May 29, 2025 ↗19% longer
is how much more time 16 experienced open source developers took with early 2025 AI tools on 246 issues in their own repositories, in METR's July 2025 randomized trial, against a forecast 24 percent speedup.
Source: METR, July 10, 2025 ↗84%
of developers were using or planning to use AI tools in their development process in the 2025 Stack Overflow survey, up from 76 percent the year before; 51 percent of professional developers used them daily.
Source: Stack Overflow, July 2025 ↗66%
of developers in the 2025 Stack Overflow survey said their biggest frustration is AI code that is almost right but not quite.
Source: Stack Overflow, July 2025 ↗48%
of 1,149 professional developers who used AI in the past year said they vibe code, and a separate 48 percent said they always check AI-assisted code before committing, in Sonar's State of Code survey published January 2026.
Source: Sonar, January 8, 2026 ↗0 to 20%
of tasks is what developers report being able to fully delegate to AI, even though they use AI in roughly 60 percent of their work, according to research by Anthropic's Societal Impacts team cited in its 2026 Agentic Coding Trends Report. The rest needs set-up, supervision, validation and human judgement.
Source: Anthropic, 2026 ↗How to vibe code well: ten practices
The directed mode, written as rules. Each one exists because of a failure that made the news.
- 01
Write the brief before the prompt
Audience, the job to be done, three to five must-have capabilities, the data the app keeps, the design direction, and what it must never do. Weak briefs are the cause of most disappointing builds, not weak models.
- 02
Ask for a plan and read it
Every serious tool now offers a plan step. Reading a plan takes two minutes and catches the misunderstanding that would otherwise cost an hour of rebuilding.
- 03
One change per round
Agents handle a single clear request far better than a list. Smaller rounds also make it obvious which change broke something.
- 04
Build in a sandbox, then promote
Separate development from production from the first day. Replit's agent deleted a live database because nothing separated them; the fix was structural, not a better prompt.
- 05
Keep secrets out of the agent's reach
API keys and passwords live in environment variables or a secrets store, never in the chat and never in a file the agent reads. Anything pasted into a chat should be treated as compromised.
- 06
Use version control even if you never read it
A commit after every working round means any bad change can be undone. Most builders and desktop apps do this automatically; check that yours does.
- 07
Test like a user, then ask the agent to test like a developer
Click through every path yourself. Then ask for automated tests on the parts that matter: payments, login, data deletion. A test the agent must keep passing is the cheapest guard rail there is.
- 08
Ask for a security review before launch
A second pass by the same or a different agent, with a checklist (authentication on every route, no secrets in code, database rules that deny by default, input validated) finds most of what the May 2025 scan found.
- 09
Record decisions in a file the agent reads
A short project file stating the stack, the rules and the reasons gives the next session the memory the last one lacked. This single habit is what separates maintainable vibe-coded apps from abandoned ones.
- 10
Know when to stop vibing
Real customer data, payments at scale, regulated information, or a codebase someone else must maintain: at that point either a developer reviews the code or the project moves to a spec-driven workflow.
What changed in 2026
This section is kept current. New entries are added as they are verified; the changelog at the foot of the page records each update.
January 21, 2026
Vibe Coding Kills Open Source
A paper by that title argued that AI-generated contributions and AI-dependent users strain the volunteer model open source runs on, which started a debate that ran through the spring.
Source: arXivFebruary 12, 2026
GitHub names an Eternal September
GitHub said open source was living through its own Eternal September, a flood of AI-generated contributions maintainers cannot keep up with, and shipped controls that let a project limit pull requests to collaborators, pin comments and set temporary interaction limits.
Source: The GitHub BlogFebruary 13, 2026
BBC shows a vibe coding platform being hacked
Security researcher Etizaz Mohsin demonstrated to BBC News that the Orchids vibe coding platform and its desktop app could be compromised, including a zero-click takeover of the reporter's laptop. The episode moved the security conversation from the apps people build to the builders themselves.
Source: BBC NewsApril 16, 2026
Microsoft Learn adds a vibe coding module
Microsoft Learn published Introduction to Vibe Coding, a nine unit training module with a graded assessment, aimed at business owners, students, startup founders and developers and built around GitHub Copilot's agent mode. The module was updated again on October 28, 2026. Codecademy lists an Intro to Vibe Coding course alongside it. The term now has a place in the official training catalogues, not only in blog posts.
Source: Microsoft LearnMay 22, 2026
The vibe slop warning
The Wall Street Journal reported developers Mario Zechner and Armin Ronacher warning of a vibe slop crisis: a volume of AI-generated code that nobody understands well enough to maintain.
Source: The Wall Street JournalMay to June 2026
rsync 3.4.3 breaks backups
The May 20 security release of the rsync backup tool shipped regressions that broke incremental backup workflows, during a period when dozens of its commits were attributed to the maintainer and Claude. Users blamed the AI co-authored changes; maintainer Andrew Tridgell disputed that and blamed gaps in test coverage. Release 3.4.4 fixed the regressions on June 8. The episode is the standard example of why generated code in infrastructure needs tests before trust.
Source: The RegisterJuly 20, 2026
New apps double, downloads do not
The New York Times reported Sensor Tower estimates that about 560,000 new apps reached Apple's App Store in the first half of 2026, close to the roughly 600,000 released in all of 2025, which was itself a 30 percent rise. Downloads grew 2 percent in the same half year, to 17.6 billion, after a 3 percent rise in 2025. Apple said it reviews 90 percent of submissions within 48 hours and holds vibe-coded apps to the same standards as any other. On October 29, 2026 The Atlantic described the result as vibe coders clogging up the App Store.
Source: 9to5Mac, reporting The New York Times and Sensor Tower, The AtlanticNovember 1, 2026
Vibe manufacturing
Business Insider reported on people using the same describe-and-iterate approach for physical objects: toys, an answering box for children, jewelry and a robot arm, prototyped with an AI model and sent to print and fabrication services. Two of the builders, Amy Zhou and Jake Fitzgerald, maintain an open source text-to-CAD tool that had passed 23,000 GitHub stars. They call the practice vibe manufacturing.
Source: Business Insider2026
Spec-driven tools become the answer to maintenance
Amazon Kiro, GitHub Spec Kit and plan-first modes in Replit Agent and Google Antigravity ask the builder to write or approve a plan before the agent writes code. Red Hat's developer team summarized the shift in February 2026: vibe anything you can validate with a test, write a spec for anything larger.
Source: Red Hat DeveloperOctober 2026
Agents get their own computers
The centre of vibe coding moved from browser builders to agent harnesses. Claude Code, Codex, Antigravity and Cursor run multi-hour builds with plans and approval gates on the builder's machine; Grok Bot, Dots and Muse run on cloud computers of their own; T3 Code and OpenCode let one person steer several of them. The practical difference from the 2025 tools is that the agent tests its own work in a browser and asks before anything sends, spends or deletes.
Vibe coding glossary
- Vibe coding
- Building software by describing it to an AI agent and accepting the generated code based on whether the result works.
- Vibe coder
- A person who builds this way; in common use, anyone shipping software with AI who does not write the code by hand.
- Vibe-coded
- Adjective for an app or codebase produced mainly by an AI agent from natural language, as in a vibe-coded landing page.
- Vibe deploying
- Extending the same approach to hosting and infrastructure: asking the agent to put the app live rather than configuring servers.
- AI-assisted engineering
- Using the same tools with full review, tests and ownership of the code. Willison's term for the professional end of the spectrum.
- Agentic coding
- Coding done by an agent that plans, uses tools, runs code and iterates on its own over many steps.
- Agent harness
- The program around the model: the chat, file access, tools, scheduler and approval gate. Claude Code, Codex, Antigravity, Grok Bot, Muse, OpenCode and OpenClaw are harnesses.
- Context engineering
- Arranging what the agent can see, such as project files, rules and examples, so that it produces the right code. The skill that replaced prompt wording as the main lever.
- Spec-driven development
- Writing a specification or plan the agent builds to, then checking the result against it.
- Plan mode
- A tool setting where the agent proposes steps and waits for approval before editing anything.
- Sandbox
- A development copy of the app with its own database where changes can be tried safely before production.
- Checkpoint
- A saved state of the project, automatic in most builders, that a session can roll back to.
- Functionality flickering
- A detail the builder never specified changing between generations because the agent re-decides it each time. GitHub Next's name for the problem specs solve.
- Productivity tax
- Time spent fixing AI output that is almost right. The top frustration in the 2025 Stack Overflow survey.
- Doom loop
- The cycle where each fix the agent makes breaks something else, usually because the builder is correcting symptoms in the running app instead of the plan or the data model. The escape is to stop, state the intended behaviour, and have the agent diagnose before it edits.
- Context rot
- The decline in an agent's output over a long session as its context window fills with old instructions, errors and dead ends. Starting a fresh session with a short written summary restores it.
- Vibe manufacturing
- The same describe-and-iterate approach applied to physical objects: an AI model generates a 3D design from a description, which goes to a printer or a fabrication service. Named in November 2026 reporting on builders prototyping toys, jewelry and robot parts.
- Vibe slop
- Large volumes of generated code nobody understands well enough to maintain, the crisis warned about in 2026.
- VibeScamming
- Using the same natural language tools to build phishing pages and fraud infrastructure.
- Hallucinated dependency
- A package the agent imports that does not exist, which attackers can register under that name. Pin dependencies and check that each one is real.
Vibe coding FAQ
What does vibe coding mean?
Vibe coding means building software by telling an AI coding agent what you want in plain language and accepting the code it produces based on whether the app works, rather than reading or writing the code yourself. Andrej Karpathy coined the term on February 2, 2025, and Collins English Dictionary named it Word of the Year for 2025.
Who invented vibe coding?
Andrej Karpathy named it in a post on X on February 2, 2025. He did not invent the practice; developers had been generating code from descriptions with ChatGPT and GitHub Copilot since 2022. He gave the all-in version a name and a permission slip for weekend projects.
Can anyone vibe code without programming experience?
Yes for a real but bounded set of projects: landing pages, internal tools, personal apps, prototypes and small products. The skills that matter are describing the outcome clearly, testing like a user and knowing when to ask for a security review. What a non-developer cannot do is judge the code itself, which is why the practices on this page exist.
Is vibe coding safe?
Not by default. Veracode found 45 percent of AI-generated code samples carried security flaws, and a 2025 scan found 170 of 1,645 Lovable apps exposing personal data. It becomes reasonably safe for small products when the builder separates development from production, keeps secrets out of the agent's reach, asks for a security pass before launch and keeps real customer data out until a developer has reviewed the app.
Is vibe coding bad?
It is bad for the things it was never meant for: production systems with real customers, regulated data, or code someone else must maintain without the agent's help. It is good for the things it was named for: fast experiments, personal software and first versions. Most criticism is of people using the pure form where the directed form, with plans, tests and review, was needed.
What is the difference between vibe coding and AI-assisted coding?
Review. In AI-assisted engineering a developer reads, tests and owns every change the AI makes. In vibe coding the builder trusts the output and judges only the result. The tools are the same; the responsibility is different.
Which tool is best for vibe coding?
For a first app with no setup, a web builder such as Lovable, Bolt or base44. For a larger project, an existing codebase or anything that must run on a schedule, a desktop agent app such as Claude Code in the Claude desktop app, the Codex app, Google Antigravity or Cursor. For work that should keep running while the laptop is closed, a cloud agent such as Grok Bot, Dots or Muse, or an open source harness such as OpenClaw or OpenCode on a server.
How do you start vibe coding?
Pick one small project you actually want, write a one-paragraph brief (who it is for, what it does, what it must not do), open a web builder, paste the brief, and use the app it produces. Paste every error back. Deploy it when it does the job. The first project should take an evening, not a week.
Does vibe coding make developers faster?
Sometimes. On new projects and boilerplate the speedup is large and obvious. On mature codebases the evidence is mixed: METR's July 2025 trial found experienced developers took 19 percent longer with early 2025 AI tools on their own repositories while feeling faster. The honest answer is that it depends on the task and on whether the developer reviews the output.
Can vibe-coded apps go to production?
Yes, and many have. The ones that work treat production as a separate environment, carry automated tests for payments and login, keep secrets in a store the agent cannot read, and have had at least one security review. Y Combinator startups and solo founders run paying products this way; the failures in the news were almost all missing one of those four things.
What are the biggest vibe coding mistakes?
A vague first prompt; building in production; pasting API keys into the chat; asking for five changes at once; never asking for a plan; skipping tests on the parts that handle money or data; and keeping no written record of the decisions, so the next session starts from zero.
Why is vibe coding controversial?
Four reasons, each with evidence behind it. Security: 45 percent of AI-generated code samples failed Veracode's security tests, and 170 of 1,645 Lovable apps exposed user data in a 2025 scan. Maintenance: code nobody has read is hard to change, which is the vibe slop warning of 2026. Productivity: the one controlled trial on experienced developers found them slower, not faster. Open source: maintainers now field AI-generated issues and pull requests from people who cannot evaluate them, and GitHub and other large projects changed their contribution rules in 2026. Some of the objection is to the name itself; Andrew Ng called it misleading because the work is still mentally demanding.
Can you vibe code with ChatGPT?
Yes, within limits. Any capable chat model, including ChatGPT, Claude and Gemini, will write a complete small program from a description, and copying that code into a file and running it is how many people first tried the practice. What a plain chat window lacks is the loop: it cannot see the project, run the app, read the error or deploy. The web builders and desktop agent apps on this page add exactly that, and OpenAI's own answer to the gap is the Codex app rather than the ChatGPT window.
Is vibe coding replacing programming?
It is replacing the typing, not the thinking. Demand for people who can specify, test and review software has gone up, and the professional end of the practice has moved toward specs and plans. Programming knowledge still decides who can take a vibe-coded app past its first few months.
Learn more on this site
Free, no signup. The beginner path runs top to bottom.
Vibe coding basics: the four step process
The beginner walkthrough of describe, generate, test, iterate.
Your first vibe coding project
A small project chosen so the first evening ends with something that works.
Choosing an AI coding assistant
A longer comparison of the builders and agent apps on this page.
Prompt engineering for coding
How to write the brief that gets a strong first pass.
Vibe coding best practices
The habits that keep a vibe-coded app maintainable.
Debugging with AI
What to paste back, and what to ask, when the build breaks.
Deploying AI-generated apps
Getting from preview to a live domain.
How to vibe code better
Briefs, tool choice and iteration for non-developers and operators.
Vibe coding launch checklist
A free pre-launch checklist covering security, data and payments.
AI agent harnesses compared
Where each agent runs, how it schedules work and what it costs.
All vibe coding guides
Webhooks, email, growth loops, scheduled routines and more.
Vibe coding articles
Dated reporting on the tools, the builders and the incidents.
Latest implementation guides
Claude Code Scheduled Tasks and /loop: Self-Checking Dev Workflows for Solo Teams
Integration · 9 minFirecrawl Skill Setup in OpenClaw: Search, Scrape, and Interact Workflows
Marketing · 8 minMarketing Agent Workflows with OpenClaw Cron, Firecrawl, and APIs
Communities
Vibe Coding is Life on Facebook has 340,000+ members sharing builds, prompts and tool news daily, and is free to join. The Vibe Coding is Life community on Skool adds courses, templates and live sessions; its free tier is open to everyone.
Sources28 references. Every number and dated event above links to one of them.
- [1]There's a new kind of coding I call vibe coding. Andrej Karpathy on X, February 2, 2025.
- [2]Not all AI-assisted programming is vibe coding (but vibe coding rocks). Simon Willison, March 19, 2025.
- [3]Vibe coding: the new way to build software. The Guardian, March 16, 2025.
- [4]Not a Coder? With A.I., Just Having an Idea Can Be Enough. The New York Times, February 27, 2025.
- [5]A quarter of startups in YC's current cohort have codebases that are almost entirely AI-generated. TechCrunch, March 6, 2025.
- [6]Vibe coding. Merriam-Webster, Slang and Trending, March 2025.
- [7]Collins Word of the Year 2025: AI meets authenticity as society shifts. Collins English Dictionary, November 6, 2025.
- [8]The hottest new vibe coding startup may be a sitting duck for hackers. Semafor, May 29, 2025.
- [9]Andrew Ng says vibe coding is an unfortunate term for an exhausting job. Business Insider, June 2025.
- [10]Vibe coding service Replit deleted a user's production database. The Register, July 21, 2025.
- [11]Measuring the impact of early-2025 AI on experienced open-source developer productivity. METR, July 10, 2025.
- [12]2025 GenAI Code Security Report. Veracode, July 30, 2025.
- [13]State of AI vs human code generation. CodeRabbit, December 17, 2025.
- [14]2025 Developer Survey: AI. Stack Overflow, July 2025.
- [15]A new worst coder has entered the chat: vibe coding without code knowledge. Stack Overflow Blog, January 2, 2026.
- [16]State of Code developer survey report: the current reality of AI coding. Sonar, January 8, 2026.
- [17]Linus Torvalds vibe coded a Python visualizer with Google Antigravity. ZDNET, January 12, 2026.
- [18]Vibe Coding Kills Open Source. arXiv, January 21, 2026.
- [19]Welcome to the Eternal September of open source: here's what we plan to do for maintainers. The GitHub Blog, February 12, 2026.
- [20]Major vibe-coding platform Orchids is easily hacked, researcher finds. BBC News, February 13, 2026.
- [21]The AI superstars who say a vibe slop crisis is coming. The Wall Street Journal, May 22, 2026.
- [22]Broken backups spark AI coding row in rsync project. The Register, June 4, 2026.
- [23]The uncomfortable truth about vibe coding. Red Hat Developer, February 17, 2026.
- [24]Introduction to vibe coding (training module). Microsoft Learn, April 16, 2026, updated October 28, 2026.
- [25]Report: App Store added nearly as many new apps in H1 2026 as in all of 2025. 9to5Mac, reporting The New York Times and Sensor Tower, July 20, 2026.
- [26]Vibe Coders Are Clogging Up the App Store. The Atlantic, October 29, 2026.
- [27]Vibe coding is old news. People are vibe manufacturing now.. Business Insider, November 1, 2026.
- [28]2026 Agentic Coding Trends Report. Anthropic, 2026.
Changelog
- 2026-11-04Weekly refresh: the 2026 timeline gains the Microsoft Learn module (April), the App Store flood (560,000 new apps in the first half of 2026 against 2 percent download growth, reported July 20 and revisited by The Atlantic on October 29) and vibe manufacturing (November 1). New statistic: developers report being able to fully delegate only 0 to 20 percent of tasks (Anthropic). Two FAQ answers added: why the practice is controversial, and whether ChatGPT can be used for it. Glossary gains doom loop, context rot and vibe manufacturing. Five sources added.
- 2026-10-30Second pass the same day: tools refreshed to the October 2026 landscape (T3 Code, Grok Bot, Dots, Muse, OpenCode, Pi, Qwen Code added; GitHub Copilot and Windsurf dropped from the table), a shorter fold, inline citation marks removed in favour of one source link per claim, source list trimmed to the references that back a number or a dated event.
- 2026-10-30Rebuilt as the full reference: definition, origin timeline, how the loop works, comparison table, two modes, tools by category, benefits, limitations with evidence, ten practices, a 2026 timeline, glossary, FAQ and sources. Previously a 700 word landing page.
First published February 9, 2026. Corrections are welcome through the about page.