{"id":71675,"date":"2026-05-13T22:18:13","date_gmt":"2026-05-13T22:18:13","guid":{"rendered":"https:\/\/www.arimetrics.com\/?post_type=encyclopedia&p=71675"},"modified":"2026-09-13T13:38:08","modified_gmt":"2026-09-13T13:38:08","slug":"vibe-coding","status":"publish","type":"encyclopedia","link":"https:\/\/www.arimetrics.com\/en\/digital-glossary\/vibe-coding","title":{"rendered":"Vibe Coding"},"content":{"rendered":"

\"VibeDefinition:<\/strong><\/p>\n

Vibe coding<\/strong> is a way of developing software through natural-language instructions, in which artificial intelligence generates the code and a person guides the result through testing and further requests.<\/p>\n

In its original sense, it describes a practice focused on observing how an application works, without necessarily reading or understanding all the generated code. The term is also used more broadly to describe programming through conversation, although this usage does not always imply the same degree of delegation.<\/p>\n\n

Origin of the term<\/h2>\n

Andrej Karpathy introduced the expression on 2 February 2025 in a post on X<\/a>. He described a development experience in which he requested changes from an assistant, ran the program and fed errors back to it, paying little attention to the implementation.<\/p>\n

Code generation using artificial intelligence already existed. The expression highlighted a particular way of using it: directing development through the program\u2019s observable behaviour, rather than writing and reviewing each modification.<\/p>\n

The term subsequently spread to other uses of conversational programming. This broader usage explains why it can refer both to experimentation without code review and, less precisely, to AI-assisted development in general.<\/p>\n

How vibe coding works<\/h2>\n

The process takes the form of an iterative conversation. A person describes the application or change they need, and the tool generates or modifies the corresponding code. After the result is run, further instructions allow errors to be corrected or its behaviour to be adjusted.<\/p>\n

A typical sequence includes these stages:<\/p>\n

    \n
  1. Description:<\/strong> the program\u2019s function, inputs and expected result are specified.<\/li>\n
  2. Generation:<\/strong> the model produces an initial implementation or modifies an existing project.<\/li>\n
  3. Execution:<\/strong> the application is tested and its behaviour observed.<\/li>\n
  4. Iteration:<\/strong> errors or desired changes are communicated through further instructions.<\/li>\n<\/ol>\n

    The prompt<\/a><\/strong> contains the instructions that guide each intervention. Some tools also receive files, screenshots or error messages as context.<\/p>\n

    For example, a person can request a budget calculator, test different amounts and ask for discounts to be incorporated or the presentation to be changed. The interaction focuses on expressing the desired behaviour and checking the result, even when modifications affect different parts of the code.<\/p>\n

    Differences from AI-assisted programming<\/h2>\n

    The main distinction is not the amount of code generated, but the person\u2019s relationship with that implementation.<\/p>\n

    In AI-assisted programming<\/strong>, a developer can delegate code writing while continuing to review the logic, architecture and changes introduced. In the original sense of vibe coding, this review takes a less prominent role than running the program and making corrections through conversation.<\/p>\n

    Tools such as Claude Code<\/a><\/strong> can be involved in either approach. Their ability to modify files or run tests does not, by itself, make the work vibe coding.<\/p>\n

    Using artificial intelligence to program is therefore not automatically equivalent to practising vibe coding. An implementation generated entirely by a model can still undergo detailed technical supervision.<\/p>\n

    Relationship with no-code and low-code<\/h2>\n

    Vibe coding, no-code and low-code reduce manual code writing in different ways, but they are not equivalent concepts.<\/p>\n

    No-code<\/strong> platforms allow applications to be configured through visual interfaces and predefined components. Low-code<\/strong> solutions combine these resources with programming to extend or adapt their functionality.<\/p>\n

    In vibe coding, the main interaction consists of asking a model to generate or modify code through natural language. The person may not read it, but the implementation still exists and determines how the application works.<\/p>\n

    These approaches can be combined: a visual platform can incorporate conversational code generation, and a project created through instructions can subsequently continue with conventional development.<\/p>\n

    Limitations of the approach<\/h2>\n

    Evaluation based mainly on visible behaviour can leave errors undetected when they arise with other inputs or conditions of use. An application can respond correctly in the cases tested while also containing flaws in its logic, permissions or data handling.<\/p>\n

    Successive modifications can also produce regressions<\/strong>: changes that solve one problem and alter a function that previously worked correctly. When the code\u2019s structure is not understood, locating the origin of these effects and maintaining the program becomes more difficult.<\/p>\n

    These limitations are not exclusive to AI-generated code. In vibe coding, they are particularly relevant because less attention is paid to the implementation. Automated tests, version control and technical review provide ways to examine changes that are not always apparent when using the interface.<\/p>\n","protected":false},"excerpt":{"rendered":"

    What vibe coding is: its origin, how it works, differences from AI-assisted programming, no-code and low-code, and limitations of the approach.<\/p>\n","protected":false},"author":33,"featured_media":79459,"template":"","encyclopedia-tag":[],"class_list":["post-71675","encyclopedia","type-encyclopedia","status-publish","has-post-thumbnail","hentry"],"_links":{"self":[{"href":"https:\/\/www.arimetrics.com\/en\/wp-json\/wp\/v2\/encyclopedia\/71675","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.arimetrics.com\/en\/wp-json\/wp\/v2\/encyclopedia"}],"about":[{"href":"https:\/\/www.arimetrics.com\/en\/wp-json\/wp\/v2\/types\/encyclopedia"}],"author":[{"embeddable":true,"href":"https:\/\/www.arimetrics.com\/en\/wp-json\/wp\/v2\/users\/33"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.arimetrics.com\/en\/wp-json\/wp\/v2\/media\/79459"}],"wp:attachment":[{"href":"https:\/\/www.arimetrics.com\/en\/wp-json\/wp\/v2\/media?parent=71675"}],"wp:term":[{"taxonomy":"encyclopedia-tag","embeddable":true,"href":"https:\/\/www.arimetrics.com\/en\/wp-json\/wp\/v2\/encyclopedia-tag?post=71675"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}