Vibe Coding Glossary: Skills, Subagents, MCP & More Explained

Plain-English definitions of vibe coding terms: Agent Skills, SKILL.md, subagents, MCP servers, AGENTS.md, hooks, plugins, plan mode and more.

Vibe coding: Building software by describing what you want to an AI model in plain language and steering by results rather than by reading every line. The term was popularised by Andrej Karpathy in February 2025.
Coding agent: An AI model running in a loop with tools (read files, edit files, run commands, search) so it can finish multi-step programming tasks on its own. Claude Code, Codex and Gemini CLI are coding agents.
Agent Skill: A folder containing a SKILL.md file (name, description, instructions) plus optional scripts and resources. The agent sees only the short description until a task needs it, then loads the rest: this is called progressive disclosure. Published as an open standard at agentskills.io.
SKILL.md: The entry file of an Agent Skill. YAML front matter holds the name and a description that tells the agent when to use it; the Markdown body holds the instructions.
Progressive disclosure: Loading context in layers: first only skill names and descriptions, then the full SKILL.md when relevant, then referenced files or scripts only when needed. It keeps the context window small even with hundreds of skills installed.
Subagent: A specialist agent the main agent can delegate to. It runs with its own context window, system prompt and tool permissions, then returns a summary. In Claude Code subagents are Markdown files in .claude/agents/.
MCP (Model Context Protocol): An open protocol, introduced by Anthropic in November 2024, that lets AI apps connect to external tool servers in a standard way. One MCP server (GitHub, Playwright, a database) works in every MCP-capable client.
MCP server: A program that exposes tools, resources or prompts over MCP. Local servers run on your machine over stdio; remote servers are reached over HTTP, often with OAuth.
MCP client: The app side of MCP: Claude Code, Cursor, VS Code with Copilot, Codex, Gemini CLI, Windsurf, Zed and many others can connect to MCP servers.
AGENTS.md: A plain Markdown file at the root of a repository with build commands, conventions and rules for coding agents. Codex, Cursor, Copilot, Gemini CLI and others read it; Claude Code reads CLAUDE.md.
CLAUDE.md: Claude Code's project memory file. It is loaded into every session in that folder and holds commands, style rules and project facts.
Plugin (Claude Code): An installable bundle of slash commands, subagents, skills, hooks and MCP servers, distributed through plugin marketplaces (git repositories with a marketplace manifest).
Hooks: User-defined shell commands that run on agent events, such as before a tool call or when the agent stops. Used for formatting, tests, notifications and guard rails.
Slash command: A saved prompt triggered by typing /name in an agent chat. Custom commands are Markdown files; plugins can ship them too.
Context window: The amount of text (tokens) a model can consider at once. Skills, subagents and compaction all exist to spend it wisely.
Compaction: Summarising the conversation so far to free context space while keeping the important facts, done automatically or with a command.
Plan mode: A read-only mode in which the agent researches and proposes a plan before it is allowed to edit files.
Spec-driven development: Writing a specification (requirements, design, tasks) first and letting the agent implement against it. GitHub Spec Kit, Kiro and OpenSpec are built around this.
Background / cloud agent: A coding agent that runs on a remote machine, works on a branch while you do something else and opens a pull request when done (Codex cloud, Jules, Copilot coding agent, Devin).
AI app builder: A browser tool that turns a prompt into a working, hosted web or mobile app with a database and auth (Lovable, Bolt, v0, Replit, Base44).
BYOK (bring your own key): Using your own model provider API key inside a tool instead of paying the tool for model usage. Common in open-source agents such as Cline, Aider and OpenCode.
Git worktree: A second working copy of the same repository on another branch. Lets several agents work in parallel without stepping on each other.
YOLO mode / auto-approve: Letting an agent run commands and edit files without asking each time. Fast, but only safe in a sandbox or container.
llms.txt: A Markdown file at a website root that gives language models a concise map of the site's documentation.
Prompt injection: Instructions hidden in content an agent reads (a web page, issue or file) that try to hijack it. The main security risk of giving agents tools; review third-party skills and MCP servers before installing.