Skip to content
Glossary Vibe Coding

Glossary Vibe Coding

All technical terms of the article The History of Vibe Coding in alphabetical order. The number at the end names the section in which the term is introduced.

A

  • Agentic engineering – way of working in which agents write the code and humans design, check and answer for it: specification, context files, tests, review, permissions; Karpathy’s counter-term to vibe coding (2026). 8
  • AGENTS.md – file in the project in which you give coding agents the rules (commands, tests, conventions, prohibitions); published by OpenAI in August 2025, with the Linux Foundation since December 2025; Anthropic’s counterpart is called CLAUDE.md. 6
  • Aider – open-source coding agent for the terminal (Paul Gauthier, 2023) that works with any model and writes its changes as Git commits. 4
  • AlphaCode – system by DeepMind (2022) that beat half the human participants in programming competitions by generating up to a million solutions and checking them against the examples. 3
  • App builder – service that builds a whole application with database and hosting from a description, without the user seeing code: Replit Agent, Bolt.new, Lovable (2024). 5
  • Artifacts – mode of Claude (June 2024) in which generated code runs next to the conversation, as a web page, chart or game; the code disappears from view. 4
  • Autocompletion – suggestion of the next name or the next lines while typing; IntelliSense (1996) from type information, TabNine (2019) and Copilot (2021) from a language model. 1
  • Automatic programming – Grace Hopper’s name (1952) for the generation of programs by programs, the first compiler; later the promise of every generation of tools. 1

B

  • Bolt.new – app builder by StackBlitz (October 2024) that reached twenty million dollars of annual revenue in two months. 5

C

  • ChatGPT – chat interface by OpenAI (November 2022) with which programmers describe problems, receive code and return error messages; made conversation the way of working. 4
  • Claude Code – coding agent by Anthropic (February 2025) that runs in the terminal, reads the project, changes files, runs commands and writes commits. 6
  • CLAUDE.md – context file for Claude Code with the rules of a project; see AGENTS.md. 6
  • COBOL – business language of 1959 whose programs were meant to read like English; billions of lines still run today. 1
  • Codex – GPT-3 further trained on code by OpenAI (2021), the basis of Copilot; since May 2025 also the name of OpenAI’s coding agent in the cloud. 3
  • Coding agent – language model with tools and a loop: plan, read files, edit, run commands, check the result, repeat until the tests pass. 5
  • Commit – saved step of change in Git with a description; agents sign themselves in as co-authors, by which their work can be recognised afterwards. 4
  • Compiler – program that translates a high-level language into machine instructions; Hopper’s A-0 (1952) and Fortran (1957) were the first, and the programmers distrusted them. 1
  • Context engineering – art of giving the model for the next step exactly the context it needs, no more and no less; term by Lütke and Karpathy (June 2025). 8
  • Context files – files such as CLAUDE.md or AGENTS.md that give an agent the rules of a project; second part of agentic engineering. 8
  • Context window – amount of text a model considers at once; grew from a few pages with Codex to whole books and determines how much of the project the agent sees. 5
  • Cursor – editor by the company Anysphere (2023), a fork of Visual Studio Code with the conversation in the editor and since November 2024 with an agent; sold to SpaceX in 2026 for sixty billion dollars. 4
  • Cyber Resilience Act – EU regulation 2024/2847: reporting duty for exploited vulnerabilities since September 2026, security requirements for products with digital elements from December 2027, regardless of who wrote the code. 7

D

  • Devin – the “first AI software engineer” presented by Cognition in March 2024: a model with terminal, editor and browser in a loop that solved 13.86 per cent of SWE-bench. 5
  • Dijkstra’s objection – text of 1978 (“On the foolishness of natural language programming”): the difficulty of programming lies in precision, not in notation. 1
  • DORA – Google’s annual study of the performance of development teams; found more speed and less stability with AI use in 2024, the “amplifier” in 2025. 4

E

  • Eternal September – GitHub’s image (February 2026) for the flood of low-quality AI contributions, after the month of 1993 in which Usenet lost its manners under new users. 7

F

  • Flash Fill – function in Excel (2013, Sumit Gulwani) that guesses the transformation of a column from two worked examples; programming by example in everyday life. 1
  • Fortran – first high-level language that caught on (IBM, 1957), announced as an “automatic coding system”; made programming many times cheaper. 1

G

  • Gemini CLI – open-source coding agent by Google for the terminal (June 2025). 6
  • GitHub Copilot – completer by GitHub based on Codex (preview June 2021, for everyone June 2022), the “AI pair programmer”; with its own agent in 2025, with fifty million users in 2026. 3
  • Glasswing – project by Anthropic (April 2026) in which selected companies check their software for vulnerabilities with the unreleased model Claude Mythos. 8

H

  • Hallucination – invented but confidently presented output of a model, such as a function or a package that does not exist. 2
  • HumanEval – measure by OpenAI (2021) with 164 programming tasks and hidden tests; counts only whether the code passes the tests. 3
  • HyperCard – program by Apple (1987) with which lay people built software from index cards, buttons and a few words. 1

I

  • In-context learning – ability of large language models (GPT-3, 2020) to solve tasks from a few examples in the prompt without being trained for them. 2
  • IntelliSense – completion by Microsoft (1996) that offers the names possible at that point while typing; precursor of the models in the editor. 1

K

  • Kiro – development environment by Amazon (July 2025) that begins with a specification the agent writes and then implements. 6

L

  • Lovable – app builder from Stockholm (November 2024), valued at 13.3 billion dollars in 2026; in 2025, 170 applications built with it were on the net with an open database. 5

M

  • METR study – controlled experiment (July 2025): sixteen experienced developers were 19 per cent slower with AI tools in their own projects and believed they were 20 per cent faster. 7
  • Model Context Protocol – open standard by Anthropic (November 2024) with which models address tools and data sources, the “USB plug” of the agents; with the Linux Foundation since December 2025. 5
  • Mythos – model by Anthropic (April 2026), not released because it found vulnerabilities in every major operating system and browser; accessible only through Glasswing. 8

N

  • n-gram model – simplest statistical language model, which counts which words follow which; beat the completion of Eclipse in 2012 and proved that code is predictable. 2
  • No Silver Bullet – essay by Frederick Brooks (1986): the difficulty of software lies in the essential, the design, not in the accidental, the tools. 1

O

  • Ollama – program with which open models run on your own computer; shows how a model continues code statistically. 2
  • Open models – language models with freely available weights, for code for instance Code Llama (2023), DeepSeek-Coder, Qwen-Coder and Kimi K2; only months behind the closed ones in 2025. 4

P

  • Pair programming – way of working from the agile methods in which two people sit at one keyboard; Copilot took over the image as “pair programmer”. 3
  • Permissions – rights an agent gets on the computer; as few as possible, with a query before commands and a sealed-off environment, because an agent is a user with powers. 8
  • Programming by example – line of research in which the user demonstrates what is wanted and the system derives the rule; Flash Fill is the most widespread example. 1
  • Prompt – input to a language model; in vibe coding the description of what the program should do, in Software 3.0 the program itself. 2
  • Pull request – proposal of a change to a project for review before it is merged; agents in the cloud deliver their work that way, and in 2026 an agent was involved in a third of all pull requests on GitHub. 5

R

  • Reasoning models – models that work through intermediate steps before answering (since o1, September 2024); improved programming markedly. 5
  • Replit Agent – app builder by Replit (September 2024); in July 2025 deleted an investor’s production database despite a change freeze. 5
  • Review – checking of a change by a second human or agent before it is merged; Amazon’s rule since 2026: two reviewers per change to critical systems. 8

S

  • Sandbox – sealed-off environment without access to the network and other people’s data, in which an agent works so that its mistakes do no damage. 8
  • Skill – procedure written down once that an agent loads when needed instead of having it explained in every conversation (Anthropic, October 2025). 8
  • Slopsquatting – attack in which somebody registers a package name invented by models and deposits malicious code; term by Seth Larson (April 2025). 7
  • Software 2.0 – Karpathy’s essay (2017): the weights of neural networks are software that no human writes but a procedure computes from data. 2
  • Software 3.0 – Karpathy’s talk (2025): prompts are programs, language models a new kind of computer, programmed in English. 8
  • Specification – description of what a program should do before code is written; at Kiro the agent writes it first, in agentic engineering a human reads it. 6
  • Stack Overflow – question-and-answer site for programmers (2008), the memory of the trade and training material of the models; lost three quarters of its questions after ChatGPT. 1
  • SWE-bench – measure from Princeton (2023) with 2,294 real bug reports from open-source projects; measures whether a model can replace a programmer. 5
  • SWE-bench Verified – cleaned subset of 500 tasks (August 2024); from 33 per cent in 2024 through 81 per cent at the end of 2025 to 97 per cent in the summer of 2026, then retired. 5

T

  • TabNine – completer that in 2019 was the first to use a language model (GPT-2); the work of a student from Waterloo. 2
  • Token – unit in which models process text and providers bill, about three to four characters; an agent consumes millions of them per task. 5
  • Transformer – architecture of neural networks (2017) on which GPT and all today’s language models rest. 2

V

  • Vibe coding – programming by telling a model what you want, not reading the code and judging only the result; word by Andrej Karpathy (2 February 2025), Collins word of the year 2025. 6
  • VisiCalc – first spreadsheet (1979), in which millions of people programmed without knowing it. 1
  • Visual Basic – language by Microsoft (1991) in which one pushed interfaces together with the mouse; the most used language of the corporate PC. 1

W

  • Windsurf – editor with agent by Codeium (November 2024); carved up between Google and Cognition in 2025 after OpenAI’s offer of three billion dollars had lapsed. 5
Last updated on