
Claude Code skill for creating and optimizing Claude Code skills. Includes optimization principles for LLM attention, router pattern, classification design, and prompt engineering.
概要
A Claude Code skill for creating and optimizing other Claude Code skills. Provides a 6-step creation process, bundled scaffolding scripts, and optimization principles for building production-quality skills. Then describe the skill you want to build. The skill walks you through understanding, planning, initializing, editing, packaging, and iterating. The skill includes an section derived from a thorough optimization pass on win4r/dev-skill. These are reusable design patterns for building high-quality skills: Position behavioral rules reference material in SKILL.md. LLMs weight earlier content more heavily — rules buried at the bottom get forgotten. When a skill has 3+ code paths and exceeds ~400 lines, split into a router (SKILL.md) + per-path files loaded on demand. Reduces initial context by 40-70%. For skills that classify user input: define a strict (resolves multi-match) and (resolves false positives). Test with 20-30 adversarial inputs.
README
以下は原文 README の抜粋です。原文を見る。
/skill-creator — Claude Code Skill Design Guide
A Claude Code skill for creating and optimizing other Claude Code skills. Provides a 6-step creation process, bundled scaffolding scripts, and optimization principles for building production-quality skills.
Usage
/skill-creator
Then describe the skill you want to build. The skill walks you through understanding, planning, initializing, editing, packaging, and iterating.
What’s Included
skill-creator/
├── SKILL.md ← Lean router: creation process + optimization principles
├── references/
│ └── skill-anatomy.md ← Background: skill structure, bundled resources, progressive disclosure
├── scripts/
│ ├── init_skill.py ← Scaffold a new skill directory
│ ├── package_skill.py ← Validate and package a skill as .zip
│ └── quick_validate.py ← Fast validation of skill structure
└── LICENSE.txt
Skill Creation Process
| Step | What | Output |
|---|---|---|
| 1. Understand | Gather concrete usage examples | Clear functionality scope |
| 2. Plan | Identify reusable scripts, references, assets | Resource list |
| 3. Initialize | Run scripts/init_skill.py |
Skill directory with template |
| 4. Edit | Write SKILL.md + bundled resources | Working skill |
| 5. Package | Run scripts/package_skill.py |
Distributable .zip |
| 6. Iterate | Test on real tasks, refine | Polished skill |
Optimization Principles
The skill includes an Optimization Principles section derived from a thorough optimization pass on win4r/dev-skill. These are reusable design patterns for building high-quality skills:
Information Architecture
Position behavioral rules before reference material in SKILL.md. LLMs weight earlier content more heavily — rules buried at the bottom get forgotten.
The Router Pattern
When a skill has 3+ code paths and exceeds ~400 lines, split into a router (SKILL.md) + per-path files loaded on demand. Reduces initial context by 40-70%.
skill-name/
├── SKILL.md ← Router: classification + shared rules
└── workflows/
├── path-a.md ← Loaded via Read tool after routing
└── path-b.md
Classification Design
For skills that classify user input: define a strict priority ordering (resolves multi-match) and disambiguation rules (resolves false positives). Test with 20-30 adversarial inputs.
Inline References
When a rule must be applied in a different section, add a reminder at the point of application. Don’t rely on the LLM to cross-reference across hundreds of lines.
Prompt Engineering Heuristics
- Keep frontmatter
descriptionunder 120 characters - Use tables for routing/lookup (O(1) for LLMs)
- Explicit “Do NOT” prohibitions in high-stakes situations
- Categorize dependencies as optional vs. required for error handling
Scaffolding Scripts
Initialize a new skill
推奨ツール
別のキーワードを試すか、フィルタを外してください。
インストール
npx skillfish add win4r/skill-creator