Claude Skills Guide

Problem: You paste the same lines into Claude Code every new session. "Follow our brand guidelines." "Use our deployment process." "Check security patterns first." Tokens burned. Minutes gone. Half the rules forgotten anyway.

Quick Win: Ship your first skill in under a minute:

.claude/skills/my-skill/SKILL.md

---
name: code-review
description: Reviews code for security and performance issues
---
 
# Code Review Process
 
1. Check for SQL injection vulnerabilities
2. Verify input validation exists
3. Confirm error handling patterns

That is a full skill. Drop the folder in your project. The next review request will pull it in automatically.

What Are Skills?

A skill is a folder of instructions that Claude finds and loads on demand. Think of a cabinet of short reference manuals, opened only when the topic fits.

Prompts vanish after one turn. MCP servers wire Claude into outside tools. Skills sit in the middle: they teach how to do things. Procedural knowledge. Workflows, checklists, house style, domain know-how.

Simon Willison put a number on it: "Each skill only takes up a few dozen extra tokens, with the full details only loaded in should the user request a task that the skill can help solve."

How Progressive Disclosure Works

Loading happens in two stages so your context window stays lean.

Stage 1 - Metadata Only (~100 tokens): Claude sees only the name and short description of each skill. Enough to spot what is on offer.

Stage 2 - Full Instructions (when needed): The body drops in the moment Claude decides a skill fits your task. That body usually stays under 5,000 tokens.

my-skill/
├── SKILL.md              # Core instructions (loads when needed)
├── scripts/              # Automation scripts
├── references/           # Documentation Claude can read
└── assets/               # Templates and files

You can carry dozens of skills this way without crowding Claude's working memory.

Skills vs Everything Else

Here is how skills line up against the rest of the stack:

Key distinction: Projects say "here's what you need to know." Skills say "here's how to do things." MCP says "here's how to connect."

When to Create a Skill

The rule: Any instruction you paste into more than one conversation should become a skill.

Good candidates for skills:

• Brand guidelines and style rules
• Code review checklists
• Deployment procedures
• Security audit patterns
• Documentation templates
• Git commit conventions

Not ideal for skills:

• One-time requests
• Project-specific context (use Projects instead)
• External API access (use MCP instead)

Writing an Effective SKILL.md

Three pieces go into every skill:

---
name: my-skill
description: What it does and WHEN to trigger it
allowed-tools: "Read,Write,Bash"
---
 
# Clear Purpose Statement
 
## Step-by-Step Instructions
 
1. First action
2. Second action
3. Expected output

Critical: The description field is what Claude scans to decide whether your skill fires. Be specific about the trigger. "When user requests code review" beats "Helps with code."

Pre-Built Skills from Anthropic

Anthropic publishes an official set of drop-in skills:

• pdf: Extract text and tables from PDF documents
• docx: Create and edit Word documents with tracked changes
• xlsx: Manipulate Excel spreadsheets
• pptx: Generate PowerPoint presentations
• brand-guidelines: Apply consistent styling

Grab them from github.com/anthropics/skills, or roll your own on the same pattern. Claude Code also ships /simplify and /batch as built-ins for reviews and codebase-wide migrations. Zero setup.

Common Mistakes

Vague descriptions: "This skill helps with documents" will never fire reliably. Name the trigger out loud.

Overloaded instructions: Your SKILL.md should stay under 5,000 words. Push long docs into reference files the skill can point at.

Missing boundaries: Say out loud what the skill shouldn't touch. That stops it firing on prompts it wasn't made for.

Your Next Steps

1. Create one skill covering the instruction you repeat most
2. Test triggering using real-world requests, not only explicit commands
3. Iterate when real sessions show what needs tightening
4. Dig into context management for more mileage per skill
5. Read up on context engineering to see where skills sit in a production AI stack

Claude shifts from generalist to specialist once skills are in place, and the context tax on repeat explanations disappears. Your expertise turns portable. For a working kit instead of a blank folder, ClaudeFast's Code Kit ships 20+ production skills across 21 categories (React patterns through database operations), plus a SkillActivationHook that picks the right skill on every prompt.

One of five moves top agentic engineers make every day. The full set lives in our Claude Code best practices guide.