Anthropic released a 32-page PDF called "The Complete Guide to Building Skills for Claude" back in late January, and it has been picking up attention as more teams start building custom Claude workflows.
A "skill" in Claude's world is a folder containing structured instructions that teach Claude how to handle a specific task consistently. Instead of re-explaining your preferred format, workflow, or coding style every conversation, you write it once as a skill and Claude applies it automatically going forward. Think of it as a persistent prompt template with actual architecture behind it.
How the Three-Layer System Works
The guide introduces what Anthropic calls "progressive disclosure" - a three-level system for loading skill information efficiently:
- Level 1 - Frontmatter: At startup, Claude reads only the name and description from each skill's YAML header. This costs roughly 50 to 100 tokens per skill, so you can install dozens without bloating the context window (the amount of text Claude can hold in memory at once).
- Level 2 - Full instructions: When Claude determines a skill matches your current task, it loads the complete SKILL.md file with all instructions, examples, and error handling.
- Level 3 - Reference files: For complex skills, additional documentation, API patterns, and templates sit in the skill folder and get pulled in only when needed.
This layered approach solves a real problem. Without it, installing ten skills would burn thousands of tokens before you even ask a question.
Who Should Read This
The guide targets four groups: developers building repeatable workflows, power users automating tasks like document generation or research, teams who want consistent Claude behavior across an organization, and builders connecting Claude to external tools through MCP (Model Context Protocol, Anthropic's standard for connecting AI models to external services).
Anthropic claims you can build and test a working skill in 15 to 30 minutes if you already know your top two or three workflows. The guide covers standalone skills, MCP-enhanced workflows, testing strategies, distribution, and common troubleshooting patterns.
The full PDF is available on Anthropic's resources site, and a companion blog post on claude.com covers the highlights. For anyone spending significant time with Claude Code or Claude for Desktop, this is the closest thing to an official curriculum on making the tool work the way you want it to.