Skip to main content

Skills Overview

Benchmarked against: Anthropic — Skills Overview Architecture: Progressive Disclosure Layer 3 Location: .claude/skills/ directory

Skills are on-demand Standard Operating Procedures (SOPs) that load into an agent's context only when invoked. They are Layer 3 of the Progressive Disclosure system — detailed procedures that would waste tokens if loaded at startup.


What is a skill?

A skill is a structured prompt file that gives an agent step-by-step instructions for a specific task. Unlike rules (always loaded), skills are loaded only when needed.

PropertyRules (L2)Skills (L3)
LoadingAlwaysOn demand
PurposeShape all behaviorGuide specific tasks
Token cost~2,000 total~500-2,000 each
InvocationAutomatic/skill-name command
ExamplesConstitution, cost awareness66s review, project scaffold

Available skills

SkillCommandPurposeToken cost
66s Review/66s-reviewSix Hats + Six Shoes systematic analysis~1,500
Create/createScaffold new project structure~800
Orchestrator/orchestratorMulti-agent communication scripts~800
Brain Lite/brain_liteEmergency quota downshift~500
Brain Mid/brain_midStandard model configuration~500
Brain Pro/brain_proFull pro capabilities~500
UI/uiFrontend component generation~1,000

How skills work

  1. User invokes with /skill-name (or system determines skill needed)
  2. Skill prompt loaded from .claude/skills/skill-name/
  3. Detailed SOP injected into current context
  4. Agent executes following the skill's step-by-step instructions
  5. Skill content persists in context until compaction

Skill directory structure

.claude/skills/
├── 66s-review/
│ └── prompt.md # The skill prompt
├── brain_lite/
│ └── prompt.md
├── brain_mid/
│ └── prompt.md
├── brain_pro/
│ └── prompt.md
├── create/
│ └── prompt.md
├── orchestrator/
│ └── prompt.md
└── ui/
└── prompt.md

Each skill is a directory containing at minimum a prompt.md file with the full SOP.


When to use skills vs. rules

QuestionAnswer → Action
Does it affect every action?Yes → Make it a rule
Is it a specific procedure?Yes → Make it a skill
Should agents always know this?Yes → Rule
Is it invoked by command?Yes → Skill
Does it have step-by-step instructions?Yes → Skill

Token economy

Loading a skill costs 500-2,000 tokens. Over a session with multiple compaction cycles, unused skills waste significant context.

ScenarioSkills loadedToken cost
Lean session (no skills)00
One review1 (66s)~1,500
Complex session2-3~2,500-4,000
All skills loaded7~5,600 (wasteful!)

Principle: Only load what you need, when you need it.


PageRelationship
Creating a SkillHow to build new skills
Design PatternsBest practices
CLAUDE.md + PDProgressive Disclosure system
Skills in CLAUDE.mdIntegration points