AI coding glossary

An AI coding assistant is a tool that uses a large language model to help you write, read and change code. It can sit in your editor, like GitHub Copilot in VS Code, or run in the terminal, like Claude Code. You describe what you want in plain words, and it suggests or makes the change for you to review.

See alsoPromptAgent

A prompt is the text you send to an AI model: your question, your instructions and any code or files you include with it. The model answers from the prompt and whatever else is already in its context. A good prompt says what you want, what done looks like and what to avoid.

See alsoContext windowRules file

The context window is the amount of text a model can work with at one time, counted in tokens. It holds your prompt, the conversation so far, rules files, file contents and tool results. When it fills up, older parts are dropped or summarised, which is one reason an assistant forgets something from early in a long session.

See alsoTokenMemory

A token is the unit a model reads and writes: a small piece of text, often part of a word. In English one token is about four characters on average. Context window sizes, usage limits and API prices are all counted in tokens.

See alsoContext window

A rules file is a Markdown file of standing instructions that an assistant reads at the start of a session: your build commands, coding style and things to avoid. Claude Code reads CLAUDE.md (and AGENTS.md when there is no CLAUDE.md), GitHub Copilot reads .github/copilot-instructions.md, and AGENTS.md is an open format that many tools read. Each tool has its own list of files and folders it looks in, so check its docs.

See alsoMemoryPrompt

SourcesAGENTS.mdClaude Code docs: memoryGitHub Docs: custom instructions for Copilot

An assistant's memory is what it carries from one session to the next, so you do not have to repeat yourself. It is saved text, such as rules files you write or notes the tool saves for itself, loaded into the context when a new session starts. The model itself does not change or learn as you work with it.

See alsoRules fileContext window

A skill is a folder of instructions that teaches an assistant one kind of task, usually a SKILL.md file with optional scripts and reference files. The assistant sees only a short description of each skill and loads the full instructions when a task needs them, which keeps the context small. Claude Code has skills, and some other tools read the same format.

See alsoPluginSub-agent

SourcesClaude Code docs: skills

A sub-agent is a separate assistant that the main one starts to handle one part of a task, with its own context window and often its own instructions and tools. It does the work, such as searching a large codebase, and sends back a short result. This keeps the main conversation focused and stops its context filling up.

See alsoAgentContext window

A hook is a command that runs on its own at a set point in an assistant's work, for example before a tool runs or after a file is edited. Hooks live in settings, not in a prompt, so they run every time instead of when the model decides to. People use them to run a formatter, block risky commands or log what happened.

See alsoSkillTool calling

A plugin is a package that adds several things to an assistant in one install, such as skills, sub-agents, hooks, commands and MCP servers. In Claude Code you install plugins from a marketplace, and each enabled plugin can add to the context of every session. A plugin runs with your permissions, so review it before you install it.

See alsoMarketplaceSkillMCP server

SourcesClaude Code docs: plugins

A plugin marketplace is a catalogue of plugins that you add to your assistant and then install plugins from. In Claude Code a marketplace is a list of plugins, often kept in a Git repository, and anyone can publish one. Check who maintains a marketplace before you add it, as you would with any code you install.

See alsoPlugin

MCP (Model Context Protocol) is an open standard for connecting AI applications to outside tools and data, introduced by Anthropic in November 2024. Instead of every assistant building its own connection to every service, both sides follow one protocol, so an assistant that supports MCP can use MCP servers built by anyone. Claude Code, GitHub Copilot and many other tools support it.

See alsoMCP serverTool calling

SourcesModel Context ProtocolAnthropic: Introducing the Model Context Protocol (25 November 2024)

An MCP server is a program that offers tools, data or ready-made prompts to an AI application over MCP. It can run on your own machine or as a remote service, and many wrap an existing API, such as GitHub or a documentation search. The assistant decides when to call its tools, and you approve each call or allow it in your settings.

See alsoMCP (Model Context Protocol)Tool calling

Tool calling is how a model takes an action outside itself, such as reading a file, running a command or searching the web. The model does not run anything: it replies with a structured request naming the tool and its inputs, and the application runs it and sends back the result. MCP is one standard way to offer tools to a model; tool calling is how the model uses them.

See alsoMCP (Model Context Protocol)Agent

An AI agent is a system that works towards a goal in a loop: it plans a step, calls a tool, looks at the result and decides what to do next, until the task is done or it needs you. A coding assistant in agent mode can edit several files, run the tests and fix what fails without a new prompt for each step. A person still has to check the result.

See alsoTool callingSub-agent

Vibe coding is building software by describing what you want to an AI and accepting the code it writes, often without reading all of it. Andrej Karpathy coined the term in a post on X on 2 February 2025. It is quick for prototypes, but code that real users will depend on still needs review, tests and a security check.

See alsoAI coding assistantHallucination

SourcesAndrej Karpathy's post on X (2 February 2025)

A hallucination is a false statement a model makes as if it were true, such as a function, package or command option that does not exist. It happens because a model predicts likely text rather than looking facts up. Running the code, reading the official docs and checking package names catch most of it.

See alsoVibe codingPrompt