Glossary

The AI coding tooling space invented a lot of new vocabulary in a short time. These are working definitions — short, neutral, with a link out to a deeper explainer when one exists.

Agent
In the broadest sense, any LLM call configured with a system prompt and a tool set. In the AI tooling space the word usually means a "sub-agent" the host can delegate to — a separate LLM call with isolated context. Browse agents
Anthropic
The company behind Claude, the LLM family, and Claude Code, the AI coding assistant. Anthropic introduced the Skill format and the Claude plugin spec.
Claude
Anthropic’s LLM family. The current generation includes Claude Opus, Sonnet, and Haiku in size order. Claude is the model that powers Claude Code.
Claude Code
Anthropic’s official CLI for working with Claude in a code project. First-party support for skills, sub-agents, MCP servers, hooks, plugins, and slash commands.
Codex
OpenAI’s coding assistant CLI. Supports skills, agents, and MCP servers, with its own on-disk conventions in `~/.codex/`.
Collection
A curated bundle of resources — typically skills, agents, hooks, and MCP configs grouped around a role or workflow. Installable as a unit; members are also installable individually. Browse collections
Cursor
A VS Code-derived AI coding editor. Supports skills, agents, MCPs, and rules; has its own on-disk conventions in `~/.cursor/`.
Front matter
YAML metadata block at the top of a Markdown file. In a `SKILL.md`, the front matter declares the skill’s name, description, and discovery flags like `user-invocable`.
Gemini CLI
Google’s command-line AI coding tool. Calls its skills "extensions"; otherwise the conceptual model is similar.
Hook
A shell command the host AI tool fires automatically on a lifecycle event — before a tool call, after a file edit, on session start. Used for enforcement and audit, not guidance. Browse hooks
Install
The act of placing a resource on disk in the right location for an AI tool to discover it on next startup. On Shared Context, "install" is one CLI command per platform.
Marketplace
A server that lists plugins (or in non-Claude tools, equivalent bundles) and points to where each one lives. Can be public or private. The Claude plugin marketplace format is a single JSON file.
MCP / Model Context Protocol
An open standard for exposing tools to LLM clients. An MCP server runs as a process and advertises tools over a JSON-RPC interface; compatible clients can call those tools during a conversation. Browse MCP servers
MCP server
A process that implements the MCP protocol and exposes a set of tools. Local or remote. Examples: Gmail, GitHub, Postgres, Slack, Linear.
Plugin
A bundle of skills, agents, MCP configs, and hooks packaged in a Claude-recognised directory layout with a `plugin.json` manifest. Installable as a single unit. How plugins work
Resource
On Shared Context: any addressable thing in the library — a skill, agent, collection, MCP config, hook, or file. The platform-neutral container.
Slash command
A user-invocable instruction the host exposes by name, prefixed with `/`. A skill with `user-invocable: true` in front matter becomes a slash command in Claude Code.
Skill
The simplest extensibility primitive — a Markdown file (typically `SKILL.md`) with YAML front matter that an AI host loads on demand and follows as a procedure. What are Claude Skills
Sub-agent
A separate LLM call delegated to by a parent agent. Has its own system prompt, tool allow-list, and context window; returns a single answer to the parent.
Sub-skill
A skill referenced from inside another skill. Useful for composing complex procedures from smaller named pieces. Discovery is a host-side convention.
System prompt
The persistent instruction set the model sees at the top of every call. Skills, agents, and the host itself contribute to this; understanding which layer adds what is core to debugging an AI tool.
Tool
A function the LLM can call. May be host-built-in (file read, shell), MCP-provided (Gmail send), or skill-described (calls another sub-agent).
Tool use
The mechanism by which an LLM requests a tool be called and receives the result. Standardised across providers as the function-calling / tool-calling pattern.
YAML
A human-readable structured-data format used for front matter on skills, agents, and plugin manifests. Whitespace-significant; if a file fails to load, check indentation first.