25 terms
The vocabulary of skills, plugins and the features around them, defined in a sentence or two. Each entry links to the guide or skill that goes deeper.
An open format for packaging instructions an AI agent loads on demand: a folder with a SKILL.md file and optional scripts, references and assets. Claude Code follows it, and several other coding agents read the same folders.
Agent Skills in other tools → See also Skill, SKILL.md
An optional frontmatter field listing tools a skill may use without asking for permission while it is active. Support varies between tools that read Agent Skills, so portable skills often leave it out.
A program shipped inside a skill, usually under scripts/, that Claude runs instead of reading. Only its output enters the context window, which makes deterministic checks cheap and repeatable.
A library for building your own agents on the same harness that powers Claude Code, available for TypeScript and Python. Agents built with it can load skills from the filesystem.
A memory file Claude Code loads into context at the start of every session, at the project or user level. Keep it short and move task-specific procedures into skills, which load only when needed.
Memory file architect skill → See also Skill
Everything the model can see at once, measured in tokens: instructions, conversation, file contents and tool output. Skills protect it by loading instructions only when a task needs them.
One or two sentences saying what the skill does and when to use it. It is the only part of a skill that sits in context all the time, so it decides whether the skill ever triggers.
Skill description tuning skill → See also Trigger
The YAML block between two lines of three dashes at the top of SKILL.md. It holds the metadata Claude reads before deciding to load a skill, such as name, description and license.
Every frontmatter field explained → See also name (frontmatter field), description (frontmatter field)
A second working directory attached to the same repository, checked out to a different branch. Running one Claude Code session per worktree lets several tasks proceed in parallel without conflicts.
Running Claude Code non-interactively from a script or CI job with the -p flag, which prints the result and exits. Skills load in headless runs the same way they do interactively.
A command Claude Code runs automatically at a lifecycle event, such as before a tool call or after a file edit. Hooks enforce rules every time, where a skill or CLAUDE.md can only ask.
An open protocol that connects AI applications to external tools and data through servers. A skill teaches Claude how to do something; an MCP server gives it new things to reach, such as a database or an issue tracker.
MCP skills → See also MCP server
A program that exposes tools, resources or prompts over the Model Context Protocol. Claude Code connects to it and can call its tools during a session.
The skill identifier: lowercase letters, numbers and hyphens, matching the folder name. It is also what you type to invoke a skill directly.
A skill installed in ~/.claude/skills/ that is available in every project on your machine. Use it for workflow habits that apply everywhere.
A package that bundles skills, commands, subagents, hooks and MCP servers so they install and update together. A plugin is described by a manifest in its .claude-plugin folder.
A catalog of plugins, defined by a marketplace.json file in a repository or at a URL. You add a marketplace once, then install any plugin it lists.
The loading model that keeps skills cheap: descriptions are always in context, the SKILL.md body loads when a request matches, and bundled files load only when a step needs them. It is why installing dozens of skills costs little context.
How skills load → See also Context window, Reference file
A skill stored in a repository's .claude/skills/ folder and committed with the code, so everyone who works in the repository gets it. Use it for stack and codebase conventions.
A Markdown file bundled in a skill, usually under references/, that SKILL.md points to for detail only some tasks need. Claude reads it only when the instructions send it there.
A folder that teaches Claude how to do one kind of task, built around a SKILL.md file. Claude reads the short description at startup and loads the rest only when a request matches.
What are Claude Code skills? → See also Progressive disclosure, SKILL.md
The required file at the root of every skill: YAML frontmatter with at least a name and a description, followed by Markdown instructions. Everything else in the folder is optional.
SKILL.md format reference → See also Frontmatter, description (frontmatter field)
A prompt you trigger yourself by typing / and a name in Claude Code. Built-in commands control the session, and custom ones run reusable prompts; current versions also let you invoke a skill directly by its name.
A separate Claude instance with its own context window, instructions and tool permissions, which the main session delegates work to. Use one when a task should not fill the main context or needs different tools.
The moment Claude decides a request matches a skill description and loads the full instructions. A skill that never triggers usually has a vague or overlapping description.