The Problem Skills Solve

Every time you start a new conversation with Claude, it starts fresh. If you always need Claude to follow a particular workflow — say, always processing PDFs a certain way, or always formatting code reviews in a specific structure — you have to re-explain that every single time.

Agent Skills solve this. They're like a permanent onboarding guide stored on a virtual filesystem — Claude reads the relevant parts automatically whenever they apply.

Skills are reusable, filesystem-based
resources that provide Claude with
domain-specific expertise: workflows,
context, and best practices that
transform general-purpose agents
into specialists. Unlike prompts
(conversation-level instructions for
one-off tasks), Skills load on-demand
and eliminate the need to repeatedly
provide the same guidance across
multiple conversations.

Three Key Benefits

Two Kinds of Skills

There are two categories of Skills, and they work identically — Claude uses them the same way regardless of origin.

Check Your Understanding

Think through each question before revealing the answer.

Three Levels of Loading

The central architectural idea in Skills is progressive disclosure. Instead of dumping everything into the context window at startup, Skills load in stages:

Level 1: The YAML Frontmatter

Every Skill begins with a YAML frontmatter block. This is the only part that is always loaded — it is what allows Claude to decide whether to invoke the Skill at all.

---
name: pdf-processing
description: Extract text and tables
  from PDF files, fill forms,
  merge documents. Use when
  working with PDF files or
  when the user mentions PDFs,
  forms, or document extraction.
---

The description field is the most important thing you write when creating a Skill. It is the sole mechanism that determines whether Claude invokes the Skill — so it must be specific about both what the Skill does and the contexts in which it should fire.

Level 3: Scripts Run Without Entering Context

The most powerful part of the Skills architecture: executable scripts that Claude runs via bash — their code never occupies the context window.

pdf-skill/
├── SKILL.md        # always read when triggered
├── FORMS.md        # read only if form-filling needed
├── REFERENCE.md    # read only if deep ref needed
└── scripts/
    └── fill_form.py  # RUNS, never read into context

The Three Levels — Flashcard Drill

Click each card to test your recall of the loading levels.

Anatomy of a SKILL.md

Every Skill requires exactly one file: SKILL.md. The SKILL.md file has two parts — the YAML frontmatter and the instruction body.

---
name: your-skill-name
description: What it does and
  when to use it.
---

# Your Skill Name

## Instructions
[Step-by-step guidance]

## Examples
[Concrete examples]

Where Skills Work

Skills are available across Claude's products, but with different constraints. The key difference is sharing scope — who can see the Skill once it's installed.

Key Facts — Fill in the Blanks

Click each blank to reveal the answer.

Apply What You Know

The Security Model

Skills run with Claude's full permissions. This is powerful — and it means you need to treat Skill sources with the same care you'd give any code you run on your machine. Four threat categories matter most.

Skills execute in the same security
context as Claude itself. Review
Skill contents before installing.
Externally sourced content may
contain prompt injection attempts.
Avoid Skills that request excessive
permissions beyond their stated
purpose. Sensitive data in Skill
context may be exposed to bundled
scripts.

Key Limitations

Skills are powerful but not universal. These are the hard constraints — not bugs, but deliberate design decisions with real practical implications.

Think It Through

These questions connect concepts across all four modules. Think before revealing each answer.

Final Quiz

Full Review Deck

Every key concept from all four modules. Shuffle and drill until they feel automatic.