What Is Vibe Coding? A Guide With Examples
Vibe coding is building software by describing what you want to an AI in plain language and accepting the code it generates, steering by results rather than reading every line. You say "make a dashboard of this month's orders with a filter by country", run what comes back, and keep prompting until it looks right. It has made software creation accessible to people who never learned to code, and much faster for people who did.
Where the term came from
AI researcher Andrej Karpathy coined "vibe coding" in early 2025 to describe fully giving in to the AI: prompting, accepting changes without close review and pasting errors back until the program works. The term spread quickly as tools like Cursor, Lovable, Bolt and Replit made the workflow mainstream.
How vibe coding works
- Describe the app or change in natural language.
- Generate: the AI writes or edits the code.
- Run it and look at the result.
- React: paste errors back or describe what to change.
- Repeat until it works.
Vibe coding examples
- A landing page for a new product, from one prompt to a deployed site.
- A personal tool, like a budget tracker or a habit app.
- A prototype to test an idea with users before engineering invests in it.
- A script to clean a CSV, rename files or call an API.
- An internal dashboard on a database, which is where the risks start to matter.
Popular vibe coding tools
| Tool type | Examples | Best for |
|---|---|---|
| AI app builders | Lovable, Bolt, Replit, v0 | Full apps from a prompt, hosted for you |
| AI code editors | Cursor, Windsurf, GitHub Copilot | Developers vibe coding inside a real codebase |
| Coding agents | Claude Code, Codex | Multi-step changes across a repository |
| AI internal tool builders | Jet Admin | Business apps on company data with permissions built in |
For a full comparison, see our list of vibe coding platforms.
Benefits
- Ideas can become prototypes quickly; time to a reliable release depends on the requirements and validation.
- Non-developers can build tools for themselves.
- Developers skip boilerplate and focus on hard problems.
- Prototypes are cheap enough to throw away.
Risks for business apps
- Security: generated code can include exposed keys, missing authorization checks or injection flaws.
- Permissions: row-level access rules are easy to describe and easy to get wrong.
- Data handling: apps may copy sensitive data into places it should not be.
- Maintainability: generated code still needs a maintainer who can explain, test, and change it.
- Ownership: vibe-coded tools become shadow IT when the builder moves on.
Taking vibe coding to production
For anything that touches company data or many users, pair the speed of vibe coding with a foundation that handles the risky parts: authentication, permissions, data connections, hosting and audit logs. Either your engineers review and harden the generated code, or you generate the app on a platform where those parts are already built. We cover this in detail in vibe coding to production and no-code vs vibe coding.
How Jet Admin fits
Jet Admin provides a platform route to generated internal tools; the app’s permissions, connected actions, and deployment still need configuration and verification. Describe the app and AI builds it on your data, across 200+ integrations queried in place, with authentication, permissions and hosting handled by the platform. Refine visually or in code with custom React, Vue or Angular components, 2-way GitHub sync and a CLI for Claude Code and Cursor. Unlimited users on every plan; granular permissions, SSO and audit logs on Business and above.
Distinguish a convincing demonstration from a maintained application
A generated screen can look complete while important behavior remains undefined. For an order dashboard, the visible chart does not establish which orders a user may retrieve, whether a refund is authorized, or what happens when a request times out. Treat the first demonstration as a way to discover requirements, then verify those requirements through the actual data and action paths.
Consider a prototype that begins with a CSV and later connects to a live order database. That change introduces credentials, record permissions, data retention, and operational ownership that the original demo may never have needed. Reassess the application when its audience or data changes instead of assuming the first successful prompt remains an adequate specification.
The OWASP AI Agent Security guidance is useful for reviewing the boundary between generated behavior, tools, and permissions. It does not make a particular app safe merely because a platform advertises agents or authentication. The deployed configuration and the user journey need their own evidence.
| Stage | Evidence to retain | Owner |
|---|---|---|
| Prototype | The intended journey and known shortcuts. | The builder who can explain the result. |
| Connected pilot | Permission tests and a representative failure. | The application and data owners. |
| Production release | A recoverable deployment and monitored outcome. | The team supporting the app. |
A practical review of a generated order tool
Begin with a small test dataset containing two customers, an invalid order state, and a failed downstream operation. Ask the generated app to show only one customer’s orders, then try retrieving the other customer’s record through a different view or request. A hidden button or filtered list is not sufficient evidence of authorization.
Review how the app validates edits and reports completion. If a save fails, the user should not see an unconditional success message. If the request times out, the app should reconcile whether the change occurred before encouraging a repeat that could duplicate an action. These are product behaviors to design, not tasks that a screenshot can prove.
Assign someone to the code, platform configuration, integrations, and recovery process. Generated code can be maintainable when a team reviews and owns it; handwritten code can be difficult to maintain when no one does. Use a small deployed change and a rollback exercise to assess the real maintenance model rather than assuming the origin of the code determines its quality.
Keep the speed benefit where it is useful: exploring layouts, trying a workflow, or producing an initial implementation. Before expanding access, record the remaining gaps and the tests that close them. This makes the transition from experimenting to operating explicit without prescribing one framework or claiming a universal delivery-time advantage.
Keep a short implementation record alongside the prototype: the current purpose, connected services, people allowed to use it, unresolved shortcuts, and the person responsible for a failure. Update that record when the app’s scope changes. This is especially useful when a successful personal experiment becomes a shared business tool, because the original prompt rarely captures all of the responsibilities that have accumulated.
Frequently asked questions
What does vibe coding mean?
Building software by describing it to an AI and accepting the generated code, guided by whether the result works rather than by reviewing each line.
Who coined the term vibe coding?
Andrej Karpathy, in early 2025.
Is vibe coding good for beginners?
It is a fast way to build personal tools and prototypes. For apps handling sensitive data, beginners should use platforms with security built in.
Can vibe-coded apps be used in production?
Yes, with review and hardening, or when generated on a platform that already provides authentication, permissions and hosting.
Vibe code your next internal tool
Start with Jet Admin for free and generate an app on your own data with security built in.