| name | skill-creator |
| description | Design and create Claude Skills using progressive disclosure principles. Use when user request to build new or update skills, or when realizing a new skill is needed, based on repeated context or domain expertise. |
Skill Creator
Principles and patterns for designing effective Claude Skills.
Quick Start
Creating a minimal skill:
- Create directory:
.claude/skills/my-skill/ - Create
SKILL.mdwith frontmatter (name, description) and body - Test in conversation
- Add references/ as content grows
When to Create a Skill
Create a skill when you notice:
- Repeating context across conversations (same schema, patterns, rules)
- Domain expertise needed repeatedly (API integration, framework conventions)
- Project-specific knowledge that Claude should know automatically
Example: "I keep explaining this database schema" → Create a database-schema skill.
Skill Architecture
Skills are directories with this structure:
my-skill/
├── SKILL.md # Required: Instructions + metadata
├── references/ # Optional: Detailed documentation
├── scripts/ # Optional: Executable operations
└── assets/ # Optional: Templates, images, files
SKILL.md - Core instructions with YAML frontmatter (name, description) + markdown body
references/ - Detailed docs loaded only when Claude needs them (schemas, API docs, guides)
scripts/ - Deterministic operations Claude can execute without generating code (validation, generation)
assets/ - Files used directly in output without loading into context (templates, images)
Progressive Disclosure
Skills load in 3 levels:
Level 1: Metadata (~100 tokens, always loaded) - YAML frontmatter determines skill triggering
Level 2: Instructions (<5k tokens, when triggered) - SKILL.md body with core patterns and links
Level 3: Resources (unlimited, as needed) - references/ scripts/ assets/ loaded only when used
Key principle: Keep Levels 1 & 2 lean. Move details to Level 3.
Development Process
- Recognize - Notice repeated context or domain needs
- Gather - Collect 3-5 concrete examples of skill usage
- Plan - Decide what goes in SKILL.md vs references/ vs scripts/
- Structure - Create directory with SKILL.md and needed subdirectories
- Write - Craft description, then SKILL.md body following patterns
- Enhance - Add references, scripts, assets as needed
- Iterate - Test in conversations, refine based on actual usage
Writing Effective Content
Descriptions: Include "Use when..." triggers, <200 chars optimal. Format: [Domain] + [Operations] + [Trigger keywords]
SKILL.md Body: Target ~50 lines, max ~150. Structure: Quick Start + Core Patterns (3-5) + Links. Use imperative voice, concrete examples.
References: Detailed docs with no size limit. Use descriptive filenames.
Common Patterns
Succeeds: Domain expertise, concrete examples, keyword-rich descriptions, clear triggers, scripts for deterministic tasks
Fails: Generic descriptions, bloated SKILL.md, second person voice, missing triggers, vague content
See skill-examples.md for detailed patterns.
Reference Documentation
For detailed guidance:
- anthropic-resources.md - Official Anthropic best practices
- writing-guide.md - Voice, structure, and code examples
Notes
- Skills are iterative - start minimal, refine through real usage
- Description drives discovery - invest time in crafting it
- Test in actual conversations to validate effectiveness
- Progressive disclosure is key - resist urge to front-load everything