Glossary

77 terms you meet while building software with AI coding agents, each explained in plain English.

.env file: A file holding your project’s settings and secrets (keys, passwords). It never goes into Git.
Absolute path: A full path starting from the root /, which means the same thing no matter where you are.
Agent team: An experimental Claude Code mode: several teammate sessions work on one task, share a task list and message each other. It is turned on with an environment variable.
AGENTS.md: A plain Markdown file of project instructions for coding agents — commands, conventions, folders to leave alone — that several tools read, so one file can serve every agent working on a repository.
AI coding agent: A tool that reads your project, plans changes, and edits real files on your machine — not just a chat window that suggests code.
API: An agreed set of rules that lets one program talk to another.
API key: A secret string that identifies your program to an online service, so the service lets it in and bills the right account. Anyone holding the key can act as you, which is why it never belongs in code.
Argument: What a command acts on — usually a file or folder name written after the command.
Branch: A parallel line of work in Git, so you can change things without touching the main version.
Claude Code: Anthropic’s coding agent: it reads your project and writes or changes code from instructions in plain language.
CLAUDE.md: A file at the root of your project holding instructions for the agent: rules, commands, conventions. It gets read automatically.
CLI: Command-line interface — driving a program by typing commands instead of clicking.
Codex: OpenAI’s coding agent, working in the same terminal-and-project shape as other agents.
Command: One instruction for the terminal, like ls or cd. You type it and press Enter.
Commit: A saved snapshot of your changes in Git, with a message saying what changed and why.
Context window: How much text the model can hold at once. When it fills up, the oldest material falls out.
Current directory: The folder you are standing in right now. pwd prints it.
Cursor: An editor with an AI agent built in — the same idea as a terminal agent, wrapped in an IDE.
Database: Organized storage you can search, add to, and update without losing track of anything.
Dependency: Someone else’s package that your project needs to run. They are listed in package.json.
Deploy: Publishing your project so other people can reach it on the internet.
Dev container: A development environment inside a container, described in .devcontainer/devcontainer.json. It gives Claude Code an isolated place to work, separate from your own system.
Diff: A side-by-side view of what changed: green lines were added, red lines were removed.
Directory: A container for files and other directories. “Folder” means the same thing.
Domain: A site’s address on the internet, like example.com — what people type in the address bar.
Dynamic workflow: A JavaScript script that Claude Code uses to run many agents: one after another, in parallel, or as a pipeline. A saved workflow runs as a slash command.
Effort: How much work Claude puts into each request: how many files it reads and how much it checks. Levels run from low to max and are set with /effort.
Environment variable: A named value programs can read, like PATH or HOME.
Executable: A file you can run as a program. The +x permission is what makes it one.
Flag: A switch that changes how a command behaves, like the -a in “ls -a”. Flags usually start with a dash.
Framework: A ready-made skeleton for applications: it handles the plumbing so you write the logic.
Frontmatter: A block of settings at the top of a Markdown file, between two --- lines. In skills and subagents it holds name, description and other fields.
Gemini CLI: Google’s coding agent for the terminal, built around very large context windows.
Git: Version control: it records the history of your code so you can see what changed and go back.
Git worktree: An extra working copy of the same repository, in its own folder and on its own branch. It lets several Claude sessions work in parallel without getting in each other’s way.
GitHub: A website that hosts Git repositories in the cloud and lets people work on them together.
Hallucination: When a model states something false with complete confidence — an API that does not exist, a function that was never written.
Headless mode (claude -p): Running Claude Code without its interactive interface: claude -p runs one prompt, prints the answer and exits. This is how it goes into scripts and CI.
Home folder: Your personal folder, written as ~. Most work starts here.
Hooks: Scripts the agent runs automatically on events — before a commit, after a file changes.
HTTP: The protocol a browser uses to ask servers for pages and data.
JSON: A plain-text format for data, written as key-value pairs. Both people and programs can read it.
Linter: A tool that automatically flags mistakes and style problems in code, like ESLint.
localhost: Your own machine’s address (127.0.0.1). A local dev server usually opens at localhost:3000.
Markdown: Formatting plain text with symbols: # for headings, * for lists. Files end in .md.
MCP: Model Context Protocol — the standard way an agent plugs into outside tools and data sources.
MVP: Minimum viable product — the smallest version that actually works and is worth showing.
Node.js: The runtime that lets JavaScript run outside a browser — on your machine or a server.
npm: The package manager for Node.js: it installs and updates the packages your project depends on.
Package: A ready-made piece of code you can install and use in your project.
Path: The address of a file or folder, like /home/user/notes.txt.
Permission mode: The Claude Code setting that decides what it does without asking: only read (Manual), edit files (Accept edits), write a plan first (Plan), or work on its own with background safety checks (Auto). Shift+Tab cycles through them.
Permissions: Who may read, change, or run a file. chmod changes them — chmod +x makes a file runnable.
Pipe: The | character, which feeds one command’s output straight into the next command.
Plan mode: A mode where the agent shows you the plan first and waits for approval — nothing changes until you say yes.
Plugin: A package for Claude Code, installed with one command, that brings skills, subagents, hooks and MCP servers. It comes from a marketplace and is enabled for you, a project, or a whole team.
Plugin marketplace: A catalog of Claude Code plugins: a repository or folder with a marketplace.json file. You add it once, then install plugins from it by a name of the form plugin@marketplace.
Process: A running program. The terminal waits for it to finish, or you stop it with Ctrl+C.
Prompt: What you ask the model. The more specific the prompt, the more useful the answer.
Prompt injection: An attack in which text an AI agent reads, such as a web page, an issue or a file, contains instructions aimed at the agent, so it acts on someone else’s words as if they were yours.
Pull request: A request to merge the changes on one branch into another, shown as a page where people read the diff, comment and approve before anything lands. GitLab calls the same thing a merge request.
Redirect: Sending a command’s output into a file: > overwrites the file, >> appends to it.
Refactor: Rewriting code so it reads better or runs cleaner, without changing what it does.
Relative path: A path read from the folder you are in right now, like ./projects or ../..
Repository: A project folder tracked by Git, together with its whole history.
Sandbox: A boundary around the commands Claude Code runs: the operating system stops them from writing outside the project or reaching network hosts that are not allowed. Turn it on with /sandbox.
Server: A program (or machine) that answers requests — handing back a web page, for instance.
Shell: The program inside the terminal that reads your commands and runs them — bash and zsh are the common ones.
Shell prompt: The short line of text before your cursor in the terminal, telling you who you are and where you are. Not the same thing as a prompt to a model.
Skill: A packaged set of instructions and tools that gives an agent a new ability on demand.
Slash command: A shortcut typed with a leading slash, like /help, that runs a predefined action.
SQL: The query language for databases: what to select, insert, or change.
Stack trace: The report a program prints when it crashes: the error, plus the chain of function calls that led to it, usually with a file name and line number for each step.
Subagent: A helper the main agent hands part of a task to, working on it in its own separate context.
Terminal: A window where you type text commands and the computer runs them.
Token: A chunk of text — roughly a word or part of one. Models measure input and output length in tokens.
Vibe coding: Building software by describing what you want to an AI agent in plain language, then steering and reviewing what it writes.