What Is Vibe Coding and What Can You Build With It?

A business owner wants a calculator that helps customers estimate a project.

A marketer needs a small interactive tool for a campaign. Someone has a repetitive spreadsheet task and wishes there were a simple app that could handle it.

Vibe coding offers a way to start building those ideas by describing them to an AI tool in ordinary language. You explain what the software should do, try the result, and ask for changes as you go.

The approach can make software experimentation more accessible. It also creates a gap that builders need to understand: generating something that appears to work is only one part of creating software people can depend on. Understanding where the term came from helps explain both its appeal and its limits.

Where the Term Vibe Coding Came From

AI researcher Andrej Karpathy introduced the term in its current AI programming sense in a post on X on February 2, 2025. He described an approach where a person could give instructions to AI and essentially “forget that the code even exists.” listserv.linguistlist.org

In that post, Karpathy discussed using Cursor Composer with a Claude Sonnet model, along with SuperWhisper for voice input. He described requesting changes conversationally, accepting the resulting edits without reading the code changes, and pasting error messages back into the tool for help. listserv.linguistlist.org

That context matters. Karpathy was describing casual experimentation, including disposable weekend projects. The distinctive feature was how far he could step away from the underlying code while still producing something functional. listserv.linguistlist.org

AI assistance for programmers already existed. GitHub introduced the technical preview of Copilot in June 2021, describing a tool that could suggest lines of code and entire functions. The arrival of the phrase “vibe coding” gave a name to a more conversational, hands-off way of working with increasingly capable coding tools. github.blog

The “vibe” captures the feel of that process. You look at what the software does, decide what feels wrong or missing, and describe the next change. The conversation might move from “make the results easier to scan” to “add a reset button” without the person manually editing the implementation.

What Vibe Coding Means Today

The phrase has acquired a broader meaning in everyday discussion. People often use it to describe building software through prompts, even when the builder also reviews the generated code.

There is a narrower interpretation, too. Developer Simon Willison argued in March 2025 that vibe coding specifically involves using a large language model to build software without reviewing the code it writes. He distinguished that practice from AI-assisted software development that includes understanding, review, and thorough testing. simonwillison.net

Both meanings are useful to recognize. In a casual conversation, “I vibe coded a dashboard” may simply mean someone used an AI tool to create it. In a technical discussion, the same phrase may imply that nobody examined how it works.

For practical purposes, the applications below describe projects you can explore through conversational AI building. As a project becomes important to a business or its customers, deliberate testing and engineering review should become part of the work.

How Vibe Coding Works

A typical session starts with a description of the desired outcome. The AI generates code, and you inspect a preview or run the program. Then you give more instructions based on what you observe.

For example, a first request might be:

Build a mobile-friendly calculator for a home cleaning business. Let visitors choose their home’s square footage, number of bathrooms, and cleaning frequency. Use clearly labeled sample pricing, show an estimated range, and explain how the estimate is calculated. Do not collect personal information.

The next requests could be much smaller:

• Put the estimate directly below the inputs.
• Explain what happens when a visitor leaves a field blank.
• Add a reset button and make sure it clears every selection.
• Show the pricing assumptions beside the result.

Depending on the tool, the AI may edit multiple files, run commands, and attempt to fix errors. Cursor’s Agent documentation describes these capabilities, illustrating how a coding assistant can act on a request rather than only return a block of suggested code. cursor.com

The person still supplies the purpose and business rules. If the calculator needs a particular pricing formula, that formula must be provided or verified. A convincing interface cannot establish whether the underlying assumptions are correct.

The finished app does not have to include AI. A calculator generated with an AI tool can run ordinary code using a fixed formula. Using AI to build software and adding AI features to software are separate decisions.

Tools Used for Conversational App Building

Several products support this general workflow, with different working environments:

When choosing a tool, consider what happens after the first preview. Can you access the code, restore an earlier version, understand the hosting costs, and get help maintaining the result? Those questions become more useful as an experiment moves toward regular use.

Practical Applications for Vibe Coding

Campaign Pages and Interactive Marketing Content

Vibe coding can help explore a landing page for a promotion, a product launch, or an event. You can describe the page structure, supply approved copy, and request variations in layout or interaction.

A useful project might be a service selector that asks visitors a few questions and points them toward a relevant offering. Another could be an interactive comparison that explains the differences between service packages.

For a marketing team, the immediate value is having something people can click through and discuss. Treat conversion improvements as a hypothesis to measure. Publishing a generated page does not establish that it will outperform the existing one.

Calculators and Estimators

Calculators are a practical starting point because their behavior can be defined clearly and checked against known examples.

Possible projects include a materials quantity calculator, a rental cost estimator, or a tool that calculates cost per lead from advertising spend and lead volume.

For the last example, you could specify that $1,000 in spend and 20 leads should produce a $50 cost per lead. You should also define what happens when the lead count is zero. Giving the AI these expectations makes the result easier to evaluate.

For customer estimates, explain assumptions and distinguish an indicative result from a binding quote.

Reporting Dashboards

A dashboard prototype can help a team decide which information it needs and how to display it. A marketing example might show campaign spend, leads, and conversion rates with filters for date and channel.

Begin with fictional or anonymized data. Ask the tool to create the charts and filters, then review the definitions behind each metric.

Connecting live reporting is a separate stage. A dashboard must obtain data from a real source and handle refreshes, missing records, and access permissions. A polished chart filled with sample numbers is useful for design review, but it does not demonstrate a working analytics integration.

Small Tools for Internal Operations

Many teams have narrow workflow problems that do not require a large software platform. Vibe coding can help explore a creative request form, a content approval tracker, or an equipment checkout tool.

For example, an agency could prototype a project intake form that requires the audience, dimensions, due date, and final approver before a request can be submitted. The team could test whether that reduces clarification emails before investing in a more permanent implementation.

Keep the initial scope specific. A useful intake form is easier to evaluate than a request to build an entire project management system.

Data Formatting and Repetitive Tasks

Vibe coding also applies to small scripts and utilities. These might standardize dates in a spreadsheet export, generate campaign tracking links, or rename a batch of image files using an agreed convention.

The opportunity comes from describing a repetitive rule clearly enough that software can carry it out consistently.

For a first experiment, work with copies and ask for a preview of proposed changes. A file renaming utility, for instance, should show the original and proposed filenames before making changes. That gives you a concrete way to check whether the instructions were understood.

Product Prototypes and Early Business Ideas

An entrepreneur can use conversational building to explore how a potential product might work. A booking flow, a directory, or a customer portal can become a clickable prototype that prospective users can try.

This helps answer focused questions. Do visitors understand the next step? Which fields are confusing? Does the proposed workflow solve the problem people described?

A successful demonstration can justify further development. Real accounts, payments, and sensitive information introduce additional requirements that a prototype may not address. Plan that transition explicitly before inviting people to rely on the product.

Educational Tools and Simple Games

Interactive learning projects provide another useful application. Examples include a vocabulary quiz, a geography game, or a training exercise that presents a scenario and explains the answer.

You can adjust the difficulty, scoring, or feedback through follow-up prompts. A trainer might test several versions of an exercise to see which one explains a concept most clearly.

The subject matter still needs review. An app can behave exactly as requested while teaching an incorrect fact or marking a reasonable answer wrong. Evaluate both the software and the educational content.

Personal Utilities

Some projects are valuable because they fit one person’s habits. A reading tracker, a hobby inventory, or a garden planning tool may only need a handful of features.

These projects can be good places to learn how to give precise instructions and test behavior. Start with a small version, such as a reading list with filters, before adding accounts or synchronization.

Also decide what should happen to saved information. If an app stores entries only in one browser, users need to understand that limitation and have an appropriate way to export anything they want to keep.

What Makes Vibe Coding Useful

The main benefit is the ability to explore an idea with less manual implementation at the beginning. A person who understands a workflow can describe it, try an early version, and discover what the original request missed.

That can improve discussions between business teams and developers. Instead of describing a complicated interaction in an email, a team can demonstrate a prototype and explain what worked during testing.

There is also a learning opportunity. Asking the AI to explain a generated function or trace how a form submission is handled can make unfamiliar programming concepts more concrete. Check those explanations against the code and documentation as your understanding develops.

The amount of time saved depends on the project. An early prototype may come together quickly while integration, debugging, and maintenance still require substantial work. Budget for the whole life of the tool when deciding whether custom software is worthwhile.

Where Vibe Coding Needs More Care

Generated software can contain incorrect logic and security problems. GitHub’s documentation explicitly notes that AI-generated suggestions can be inaccurate or insecure and should be carefully reviewed and tested. docs.github.com

For a business application, examine who can access each record, where information is stored, and whether credentials are protected. Check that important calculations use approved rules. Decide how the application will be monitored and maintained after launch.

These concerns become especially significant when software handles customer accounts, takes payments, or supports decisions with serious consequences. Bring in someone qualified to evaluate the relevant parts before relying on the result.

Maintenance also deserves attention. Each new feature adds behavior that future changes can break. If nobody understands the structure of the application, a small request may become difficult to diagnose. Keep the code and its documentation available to whoever will support it.

How to Start With a Manageable Project

Define One Outcome

Choose a task with a clear result. “Build a tool that creates campaign tracking URLs from five inputs” gives you something specific to test. “Build our marketing software” leaves too many decisions unresolved.

Write down who will use it, the inputs they provide, and the output they should receive.

Give the AI the Rules

Provide formulas, approved text, examples, and constraints. Say whether the project needs to store data or connect to another service. If you want a temporary prototype with sample information, state that directly.

Ask the tool to identify missing requirements before building. Its questions can help expose assumptions, although you should still review the plan yourself.

Build and Check One Feature at a Time

Start with the central interaction. Test that before adding secondary features. Make changes in small enough steps that you can see which request caused a problem.

Save working versions so you can return to them if an experiment breaks the project.

Test Beyond the Ideal Example

Try blank fields, unusually large values, repeated clicks, and the browsers or devices your audience uses. For a form, confirm what happens when submission fails. For a calculator, compare several results with independent calculations.

Also check keyboard navigation and whether labels and error messages make sense without relying on color alone. These checks help reveal problems a visual preview can miss.

Decide What Must Happen Before Launch

List the unfinished work honestly. A prototype might still need accurate production data, accessibility improvements, technical review, or a hosting and maintenance plan.

Assign responsibility for those tasks. If the tool will support an important business process, someone needs to own its ongoing reliability.

Choosing Your First Vibe Coding Project

Look for a small problem you understand well enough to judge the result. A calculator with a known formula, a campaign planning utility, or a simple interactive explanation gives you a useful place to begin.

Describe the outcome, supply realistic examples, and test the behavior yourself. As the project grows, use what you learn to decide whether to keep it as a personal experiment, improve it with a developer, or move the idea into an existing platform. The most valuable result is a tool that serves a clear purpose and can be supported at the level its users need.