AI Glossary / Generative AI and language / Advanced
Vibe coding
Definition
Vibe coding is an informal term for creating software by describing desired behavior to a generative AI system and iterating on what it produces. Its original, narrower meaning emphasized accepting generated code with little attention to its internal workings. Broader usage now sometimes includes AI-assisted programming generally. Because these meanings differ, an account of a project should explain the actual development and review process rather than rely on the label alone.
What it means
A person might request a small application, try the result, describe what feels wrong and ask the AI to revise it. Much of the interaction takes place through ordinary language rather than manual editing of source code.
The important distinction is how the generated result is evaluated and understood. Exploring a disposable prototype through its visible behavior differs from maintaining software whose implementation has been inspected and tested.
The term was coined by Andrej Karpathy in 2025. Google's glossary explicitly notes the shift from an intentionally loose original meaning to broader use.
How it works
The conversational loop typically involves stating a goal, generating code, running or previewing it, observing the result and requesting changes. Some tools can edit files and run checks; others return code that the user must place and execute elsewhere.
A prompt is still a specification, even when informal. Missing details can lead the system to choose behavior the user did not intend. Feedback helps refine the result, but a successful demonstration covers only the situations actually exercised.
As a project becomes more consequential, the relevant question is whether its behavior and implementation have been adequately checked.
Three practical examples
- A workshop participant describes a browser-based vocabulary game and adjusts its layout through follow-up prompts. The prototype helps the group decide whether the activity is engaging.
- A hobbyist creates a local timer with sound and color changes. Testing includes resetting, pausing and reopening it, rather than only watching one normal countdown.
- A product team produces a mock scheduling screen for a discussion. Before turning it into an operational service, the team investigates how dates, saved changes and simultaneous edits actually behave.
Think of it this way
Imagine shaping a rough model by repeatedly telling a collaborator what to change.
However, a convincing exterior does not reveal every internal assumption. When the model becomes something other people depend on, its hidden behavior matters alongside how it looks and feels.
Worked example
Consider an original teaching exercise: a user asks for a reading-progress calculator. The intended percentage is completed pages divided by total pages, multiplied by 100.
With 30 completed pages out of 120, the expected result is 30/120 × 100 = 25%. A generated calculator displays 25%, so the ordinary demonstration succeeds.
The user then tries 0 out of 120, expecting 0%; 120 out of 120, expecting 100%; and 30 out of 0, which needs a defined response because division by zero is not a meaningful progress percentage. Another input, 150 out of 120, raises a product decision: reject it, allow progress above 100%, or handle it in another explicitly chosen way.
Suppose the first three cases pass and the last two reveal missing behavior. Three passes out of five listed cases is 60% of that small test set. It is not evidence that the application is 60% correct overall. The cases are deliberately selected examples, not a representative probability sample.
A useful revision prompt states the chosen rules and expected outputs. Merely asking the AI to “make it work better” leaves those decisions implicit.
Illustrative case study
Fictional Cedar Workshops creates a sign-up prototype by describing screens to an AI assistant. In a meeting, adding and removing attendees appears to work, so the team initially assumes the application is nearly finished.
A later trial with two browser sessions reveals that one person's changes can overwrite another's. The visible single-user demonstration never tested that situation. The team documents the required shared-update behavior and has the implementation revised and reviewed before using it for actual bookings.
The prototype still served a purpose: it made the desired workflow easier to discuss. The mistake was treating that evidence as proof of readiness for a different operating context. The team now distinguishes exploration milestones from acceptance criteria for the working service.
Common misconceptions and mistakes
AI-generated code is still code. Its dependencies, data handling and maintenance needs do not disappear because the instructions were conversational.
Using an AI assistant does not necessarily imply the narrow, low-inspection meaning of vibe coding. A developer may review every change and understand the resulting implementation.
A polished interface is not a complete quality assessment. Behavior under unusual inputs or concurrent activity may differ from the demonstration.
More prompting is not always the right remedy. Sometimes the next useful step is clarifying a requirement, examining the implementation or testing a specific assumption.
Business use
Vibe-coding workflows can help teams explore ideas and communicate desired behavior. Evaluate that benefit against the complete effort required to produce and maintain the intended result.
Record time spent on generation, clarification, debugging, review and later corrections. A fast first screen is only one part of that total.
For a shared application, identify who owns the requirements and who can assess the implementation. Keep the actual checks visible in project records.
People also ask
Is vibe coding the same as all AI-assisted programming?
Usage varies. The original narrow meaning and broader contemporary usage should be distinguished.
Do I need to type source code myself?
Some workflows permit mostly natural-language interaction, but someone still needs to assess whether the result meets its intended requirements.
Does a working preview mean the application is finished?
No. It demonstrates the behavior exercised in that preview, not every relevant condition.
Check your understanding
Why is 30 out of 120 a limited test?
Answer: It checks one ordinary input and misses boundary behavior.
Who decides what 150 out of 120 should mean?
Answer: The application owner or requirements process, rather than an unstated assumption.
Why describe review practices explicitly?
Answer: The term vibe coding alone does not establish how thoroughly the software was checked.
Related terms
- Conversational coding
- Prompt
- Software testing
- AI-assisted programming
- prototype
- code generation
- specification
- code review
Learn next
Choose a small prototype and write five expected behaviors before requesting revisions. Include an ordinary case, an empty input and a boundary case. Compare the observed results with those expectations and record unresolved product decisions separately from implementation defects.
More in Generative AI and language
- Unidirectional language model - A unidirectional language model predicts tokens using context from one direction in a sequence.
- Unsupervised Pretrained Network - An unsupervised pretrained network is a neural network whose parameters are first learned using an objective that does not require the eventual task's...
- Variational autoencoder - A variational autoencoder, or VAE, is a probabilistic generative model that learns a latent-variable representation of data together with an approximate...
- Vector quantization - Vector quantization represents an input vector by one of a finite set of representative vectors called a codebook.
- Word tokenization - Word tokenization divides text into word-level units and, depending on the rules, separate punctuation or other tokens.
- Zero Redundancy Optimizer - Zero Redundancy Optimizer, usually written ZeRO, is a family of distributed-training memory optimizations that partitions model states across data-parallel...
Sources
Google, Machine Learning Glossary: Generative AI, vibe coding and conversational coding definitions, including the term's changing usage. https://developers.google.com/machine-learning/glossary/generative Google Cloud, What is vibe coding?, conversational workflow and distinction between exploratory use and reviewed AI-assisted development. https://cloud.google.com/discover/what-is-vibe-coding The calculator and fictional workshop case are original teaching illustrations. The account of the term's origin follows Google's description; the original social post was not directly retrievable.