Anthropic Skills Explained: Fancy Prompts or Something Actually Useful?

Cutting through the jargon to understand what Skills actually are.

I'll be honest: when I first heard "Skills," my eyes rolled.

We already have:

• Agents (sub-processes that do tasks)
• Slash commands (shortcuts to prompts)
• Custom instructions (persistent context)
• CLAUDE.md (project memory)
• MCP servers (external tool connections)

And now... Skills?

Is this just another fancy name for "a prompt you saved somewhere"? Is Anthropic inventing vocabulary to sound more sophisticated?

I spent a week figuring this out. Here's what Skills actually are, when they matter, and whether you should care.

Table of Contents

1. What Are Skills, Really?
2. Skills vs. Everything Else
3. The Progressive Disclosure Architecture
4. Real Examples: Skills in Action
5. How to Create Your Own Skill
6. Where Do Skills Work?
7. Is Anthropic Watching Your Skills?
8. The Agent Skills Standard
9. Should You Actually Use Skills?

1. What Are Skills, Really?

Let me give you the honest definition:

Skills are folders containing instructions and scripts that Claude loads on-demand to perform specific tasks.

That's it. A folder with a SKILL.md file and optionally some code.

But here's what makes them different from "just a prompt":

Skills Can Include Code

A skill isn't just text instructions. It can bundle Python scripts, reference files, templates, and resources. When Claude uses the skill, it can execute that code.

Example: The PDF skill doesn't just tell Claude "here's how PDFs work." It includes a Python script that extracts form fields from PDFs. Claude runs the script, gets structured data, and works with that.

Skills Are Auto-Invoked

You don't type /pdf-skill. You just say "extract the form fields from this PDF" and Claude automatically recognizes it needs the PDF skill, loads it, and uses it.

This is the key difference from slash commands: Claude decides when to use skills based on your request.

Skills Are Portable

The same skill works in:

• Claude.ai (web)
• Claude Desktop
• Claude Code
• Claude API

Write once, use everywhere.

2. Skills vs. Everything Else

Let me clarify the confusion:

The Key Insight

Skills = Instructions + Code + Auto-invocation + Portability

Slash commands are explicit. You decide when to use them. Skills are implicit. Claude decides when to load them.

That's the fundamental difference.

3. The Progressive Disclosure Architecture

This is where Skills get clever.

Anthropic designed a three-tier system:

Tier 1: Metadata (always loaded)
├── Skill name
└── Skill description

Tier 2: Core Instructions (loaded when relevant)
└── Full SKILL.md content

Tier 3: Supplementary Resources (loaded on-demand)
├── reference.md
├── templates/
└── scripts/

How It Works

1. At startup, Claude loads ONLY the names and descriptions of all installed skills
2. When you make a request, Claude checks if any skill description matches
3. If relevant, Claude loads the full SKILL.md into context
4. If needed, Claude loads additional files from the skill folder

This is like a well-organized manual:

• Table of contents first (fast lookup)
• Specific chapters when needed (relevant detail)
• Appendices on-demand (deep resources)

Why This Matters

Context windows are finite. Loading every skill fully would waste tokens.

Progressive disclosure means skills can be effectively unlimited in size while only loading what's needed for each task.

4. Real Examples: Skills in Action

Let's look at actual skills from Anthropic's repository:

PDF Skill

Purpose: Extract and fill PDF form fields

What it includes:

• Instructions for identifying form fields
• Python script for PDF manipulation
• Guidelines for common PDF operations

How you use it:

"Extract all form fields from contract.pdf and fill in the company name as 'Acme Corp'"

Claude:

1. Recognizes this is a PDF task
2. Loads the PDF skill
3. Runs the Python script to extract fields
4. Fills the values
5. Returns the modified PDF

Without the skill: Claude would describe how PDFs work conceptually but couldn't actually manipulate them.

Brand Guidelines Skill

Purpose: Apply company brand standards to content

What it includes:

• Color palette specifications
• Typography rules
• Voice and tone guidelines
• Do's and don'ts

How you use it:

"Write a blog post about our new product launch"

Claude:

1. Recognizes content creation task
2. Loads brand guidelines skill
3. Applies brand voice, approved terminology
4. Formats according to brand standards

Without the skill: Claude writes generic content that needs manual brand alignment.

MCP Builder Skill

Purpose: Create Model Context Protocol servers

What it includes:

• MCP specification reference
• Code templates for common server patterns
• Testing procedures
• Deployment guidelines

How you use it:

"Create an MCP server that connects to our internal API"

Claude:

1. Loads MCP builder skill
2. Uses templates and patterns
3. Generates properly structured server code
4. Includes correct protocol implementation

Without the skill: Claude might generate MCP code with subtle protocol errors.

All 16 Official Skills

5. How to Create Your Own Skill

Minimum Structure

my-skill/
└── SKILL.md

SKILL.md Format

---
name: my-custom-skill
description: A clear description of what this skill does and when Claude should use it
---

# My Custom Skill

## Purpose
Explain what this skill helps accomplish.

## Instructions
Step-by-step guidance for Claude.

## Examples
Show input/output examples.

## Guidelines
- Do this
- Don't do that

With Code

my-skill/
├── SKILL.md
├── scripts/
│   └── process.py
└── templates/
    └── output-template.md

In your SKILL.md:

## Execution
When processing data, use the script at `scripts/process.py`.

Installation in Claude Code

# Add skill to project
mkdir -p .claude/skills/my-skill
# Add your SKILL.md and resources

# Or install from marketplace
/plugin marketplace add anthropics/skills

Pro Tips

1. Name matters: Claude uses the name to match tasks to skills
2. Description is critical: Be specific about when to use this skill
3. Keep SKILL.md focused: Put supplementary info in separate files
4. Include examples: Claude learns from examples in the skill

6. Where Do Skills Work?

Cross-Platform Portability

This is the big win. A skill you create works everywhere:

~/.claude/skills/my-skill/
└── Available in Claude.ai, Desktop, AND Code

Slash commands are Claude Code only. Skills are universal.

7. Is Anthropic Watching Your Skills?

Short answer: No, not by default.

Telemetry Details

Claude Code supports OpenTelemetry for monitoring, but:

• It's opt-in (not enabled by default)
• Sensitive data is never included
• Prompt content is redacted by default
• You must explicitly set OTEL_LOG_USER_PROMPTS=1 to log prompts

What Anthropic DOES See

If you use Claude.ai or the API:

• Standard usage metrics (token counts, session data)
• Error reports for debugging
• Model feedback when you thumbs-up/down responses

What Anthropic DOESN'T See

• Your custom skill contents (unless you publish them)
• Your prompt content (unless you opt-in)
• Files processed by skills

Skills Are Local

Your skills live in:

• ~/.claude/skills/ (global)
• project/.claude/skills/ (project-specific)

These are just folders on your machine. Anthropic doesn't have access to them.

8. The Agent Skills Standard

Here's something interesting: Anthropic made Skills an open standard.

What This Means

The specification at agentskills.io defines:

• How skills are structured
• How agents discover and load them
• How cross-platform compatibility works

Adopters

Skills aren't just for Claude:

• Cursor supports skills
• VS Code (via extensions)
• GitHub Copilot (planned)
• Gemini CLI (community support)
• OpenCode (community support)

Why Anthropic Open-Sourced It

From their engineering blog:

"Rather than building fragmented, custom agents for each task, organizations can now specialize general-purpose agents by packaging procedural knowledge into reusable capabilities."

They want skills to become the standard way agents learn domain expertise—not just for Claude, but for all AI agents.

The Vision

Long-term, Anthropic envisions agents that can:

• Create their own skills
• Edit and improve skills
• Share skills with other agents
• Evaluate skill effectiveness

Skills as self-evolving agent capabilities.

9. Should You Actually Use Skills?

Yes, If:

No, If:

The Honest Reality

One reviewer noted:

"I've struggled to get Claude to invoke a skill without me explicitly asking it to."

Auto-invocation isn't perfect. Sometimes you need to say "use the X skill" anyway.

But when it works, it's seamless. And the cross-platform portability is genuinely useful.

My Take

Skills are not just fancy prompts.

They're:

• Portable (work everywhere Claude works)
• Composable (instructions + code + resources)
• Automatic (Claude loads them when relevant)
• Standardized (open spec, multi-platform)

Are they revolutionary? Not quite. They're an evolution of prompt management into something more structured and capable.

Should you invest time in creating skills? Only if you have repeatable workflows that would benefit from bundled code and cross-platform use.

For most developers, slash commands in Claude Code are simpler and explicit. But if you're building organizational workflows or need the same capabilities across Claude.ai and Claude Code, skills are worth learning.

Written while testing the skill-creator skill to create a skill about creating skills. It's skills all the way down.

Sources:

• GitHub: anthropics/skills
• Agent Skills Standard
• Anthropic Engineering: Agent Skills
• Claude Skills Documentation
• ProductTalk: Claude Code Features