Vibe Coding Glossary: Key Terms in Plain English

In short

Vibe coding has grown its own jargon: agents, specs, MCP, tokens, hallucinations. If half the words in the articles and videos sound like incantations, this glossary is for you. It covers 30 key terms, each explained in one or two plain sentences, with links to detailed breakdowns for most of them.

You don't have to read it front to back — use it as a cheat sheet. The terms are grouped: the basics first, then how models work, then prompts and specs, then agents and tools, and finally the model's knowledge (RAG and fine-tuning). The term "vibe coding" itself was coined by Andrej Karpathy (former director of AI at Tesla) in February 2025, for the approach where you describe the task in words and the AI writes the code.

What is vibe coding, in plain language?

Vibe coding is a way of building software where you state the task in ordinary language and a neural network writes and fixes the code for you. Covered in detail in what vibe coding is.

Vibecode is the colloquial name for both the approach and the result: "threw together a vibecode in one evening". Often with a hint of "it works, but I'm scared to look under the hood".

Vibe coder is a person who builds products through a dialogue with AI rather than by writing every line by hand. Who they are and what they can do is in our profile of the vibe coder.

No-code is building applications with no code at all, in visual builders (blocks, buttons, a mouse). Unlike vibe coding, there's no source code here that you can take away and extend.

Low-code is an approach where most of the work is assembled visually, but you can still write code by hand for the tricky parts.

The main difference between vibe coding and classic development is worth grasping right away: here you steer the result through text in ordinary language instead of typing every line of code yourself.

What are an LLM, a token and context?

LLM (large language model) is a neural network trained on huge volumes of text that can continue it meaningfully. ChatGPT, Claude, Gemini and other assistants run on LLMs.

Neural network is a mathematical model that learns to find patterns in data from examples rather than from hard-coded rules.

Token is the smallest chunk of text a model operates on: part of a word or a few characters. Tokens measure both the length of a request and the cost of API usage.

Context (context window) is the amount of text the model "holds in its head" within one dialogue, measured in tokens. Anything that falls outside the window is forgotten.

Hallucination is a confident but invented answer from the model: a function, link or fact that doesn't exist. Taking hallucinations on faith is the main risk of vibe coding.

Inference is the moment a finished model produces an answer to your request (as opposed to training, which happened beforehand).

Multimodality is a model's ability to work not only with text but also with images, audio and code in a single request.

Prompt, spec and context engineering — what's the difference?

Prompt is your text request to the model: what to do and how. Prompt quality directly affects the result. How to write them and which templates to build on is in our complete guide to prompts.

System prompt is a persistent instruction that sets the model's role and rules for the whole dialogue ("you are a senior Python developer").

Prompt engineering is the skill of phrasing requests so you get the result you need consistently, not on the tenth attempt.

Spec (specification) is a detailed brief for the AI: what the product is, which screens, what logic and constraints. In serious vibe coding, a spec replaces chaotic prompts and keeps the agent within bounds.

Context engineering is managing what the model sees in its context: which files, examples and rules to hand it so it doesn't have to guess.

Few-shot (learning from examples) is the technique of giving the model 2–3 "input → output" examples right in the request so it copies the pattern.

What is an AI agent and what is MCP?

Agent (AI agent) is an LLM-based program that doesn't just answer but carries out steps itself: reads files, writes code, runs commands, checks the result. Claude Code and Cursor are examples of coding agents.

Agentic loop is the repeating circle of "think → take an action → look at the result → adjust" until the task is solved.

Subagent is a separate agent the main one launches for a specific subtask (run the tests, work through one module) with its own separate context.

MCP (Model Context Protocol) is an open standard through which an AI agent connects to external data and tools: databases, files, APIs. Anthropic released it in November 2024; it's often compared to "USB-C for AI applications". How to connect it is in the MCP guide for Claude Code, and ready-made servers are in the MCP arsenal.

Skill is a packaged instruction the agent loads for a particular type of task (for example, "how to format reports"). How a skill differs from MCP and an API is in a separate breakdown.

Tool is a specific function the agent can call: read a file, go to the internet, run a test.

Hook is an "on event" rule: for example, automatically run the linter after every code edit.

RAG and fine-tuning — how do they differ?

RAG (retrieval-augmented generation) is an approach where the model pulls the relevant documents from your knowledge base before answering and responds based on them rather than on memory alone. It reduces hallucinations.

Embedding is a numeric representation of text that lets a machine find fragments similar in meaning. The basis of retrieval in RAG.

Vector database is a store of embeddings where search works by meaning rather than by exact wording.

Fine-tuning is additional training of an existing model on your own data so it handles a narrow task better. More expensive and more complex than RAG, and needed less often than people think.

API is a way to call the model programmatically, from your own code, instead of through a chat window.

Cheat sheet: what gets confused with what

FAQ

How many terms does a beginner really need?

Five will do: prompt, context, agent, spec and hallucination. With those you already follow most articles and videos about vibe coding, and the rest comes with practice.

How is a skill different from MCP?

A skill is knowledge ("how to do it") that the agent reads and applies. MCP is the channel to data and tools ("where to get it and what to act with"). A skill can use tools connected through MCP. More detail in our breakdown of MCP, skills and APIs.

Do you have to know how to program for vibe coding?

To get started, no: you really can assemble a basic product through a dialogue with AI. But understanding the fundamentals helps you spot hallucinations and avoid breaking the project out of nowhere. And it's worth at least recognizing basic web syntax on sight — that makes it easier to read what the AI generated and catch mistakes in time.

Where can I apply these terms in practice?

In qvib's Quest engine, specs, MCP servers, skills and subagents are already assembled and configured for Claude Code and Cursor — no need to glue it all together by hand. See how the engine works on the engine page, and the learning materials are collected in the knowledge base hub.