Building software by describing what you want to an AI agent and shipping the result, often without reading every line of code that runs.
Research-based overview. This article synthesizes public documentation, pricing pages, and user reports. How we research.
Vibe coding is a 2025 coinage describing a way of writing software where the developer prompts an AI in natural language, accepts most of the AI's output without line-by-line review, and judges correctness by whether the program runs and feels right rather than by reading the source. The term was popularized by Andrej Karpathy in a February 2025 tweet:
“There's a new kind of coding I call ‘vibe coding', where you fully give in to the vibes, embrace exponentials, and forget that the code even exists. It's possible because the LLMs (e.g. Cursor Composer w/ Sonnet) are getting too good. Also I just talk to Composer with SuperWhisper so I barely even touch the keyboard.”
The phrase took off because it named something many developers had quietly been doing for months but did not have a label for. Within weeks, "vibe coding" had spread across X, Hacker News, and product launches. By mid-2025, tools like Lovable, Bolt, and Replit Agent were marketing directly to vibe coders. By 2026, the term has settled into the working vocabulary of indie software builders, AI-native developers, and a growing share of mainstream engineers. There is also a backlash — some developers reject the term as glorifying the abandonment of code review — which is itself a sign that the practice has reached critical mass.
The pattern is loose by design, but it generally moves through four stages:
You describe what you want in plain language. Not a spec, not pseudocode — closer to a wish. "Build a Pomodoro timer with a leaderboard of who has the longest streak." "Add a settings page where users can change their email." "Make it look like Linear." The prompt is short, declarative, and assumes the AI fills in the rest.
The AI produces working code. In Cursor, this might be a multi-file diff. In Lovable or Bolt, a full app boots up in a sandbox. In Claude Code or similar terminal-based agents, files are written and edited across the project. The output runs without you typing the implementation.
You run the result. If something is wrong — visually, behaviorally, or logically — you describe the problem and the AI patches it. The loop is short, often dozens of iterations per hour. You spend most of your time judging output, not authoring it. Crucially, you are usually not reading the source between iterations.
Once the program looks right, you ship. For a one-off tool, this might be a Vercel deployment. For a SaaS, it goes to a small set of users and you continue iterating in production. The loop never really closes — new prompts produce new features, indefinitely.
The tooling has fragmented into rough categories. The same product often slots into multiple categories.
These show you a running app from prompt one and let you iterate in the browser. Best for marketing pages, MVPs, and small SaaS products.
You retain a code editor and the agent works inside it. Best when you have an existing codebase or want to stay close to the source.
You describe what you want and the agent executes shell commands, edits files, and runs tests. Best for backend, devops, and existing codebases.
For a deeper categorization with strengths and weaknesses, see our roundup of vibe coding tools. For a head-to-head between the three most-used IDEs and builders, our Cursor vs Lovable vs Claude comparison covers when each one wins.
| Traditional coding | Vibe coding | |
|---|---|---|
| Author of source | You | AI agent |
| Primary unit of work | The function | The prompt |
| How you verify correctness | Reading code, writing tests | Running the program, observing behavior |
| Where bugs surface | At review or test time | At runtime, often in production |
| Speed of MVP | Days to weeks | Hours to days |
| Quality at scale | Good with discipline | Degrades past ~10k LOC without intervention |
| Onboarding cost for collaborators | High — they read the code | Higher — the code may have no consistent author or pattern |
| Best fit | Long-lived production systems, libraries, infrastructure | Prototypes, internal tools, MVP SaaS, single-founder products |
The honest list. Vibe coding is real and useful but breaks down in predictable ways.
None of this disqualifies vibe coding. It just defines where it works. For solo founders shipping MVPs and small SaaS, the speed advantage is overwhelming. For teams maintaining systems that handle credit cards or healthcare data, vibe coding is at best a starting draft. Our walkthrough of how to build a SaaS with Claude shows where the practice fits naturally and where you should slow down and read the diff.
"Vibe coding" will probably not be the term we use in 2028. It is too informal to survive in enterprise contexts, and the practices it describes are already splintering — agentic coding, AI-pair programming, prompt-driven development, autonomous engineering — each with slightly different connotations. What will persist is the underlying shift: the unit of work in software development is moving from the function to the intent, and from the keystroke to the prompt. Whether we keep calling it vibe coding or rename it for the next era, the loop of prompt → generate → iterate → deploy is now part of how software gets made, and that change appears permanent.
The stack, prompts, pricing, and mistakes to avoid — for solo founders building with AI.