create-expert-skill

Transform expert conversations into production-grade Claude Skills. Whith this enhanced "skill-creator" skill you can capture domain knowledge and system-specific ontologies through structured roleplay. It also packages deterministic scripts alongside flexible guidance, and loads expertise progressively, turning AI assistants into specialists.

Why This Exists

Anthropic released a basic "skill-creator", however, it doesn't utilize the entire range of what's possible within a Skill. This enhanced skill creator makes use of resources, examples, templates, scripts, progressive disclosure and system architecture knowledge to deliver elaborate skills, zipped and ready to upload.

Why the "Expert" Part Matters

AI assistants struggle in production for two reasons:

1. Missing domain expertise — Generic models don't know or aren't primed with your industry's edge cases, terminology, or unwritten rules.
2. Missing ontology understanding — They don't grasp your specific data structures, entity relationships, or system constraints

This skill solves both by helping you:

• Interview experts (or yourself) to extract implicit domain knowledge
• Map system ontologies — company-specific structures, codes, and relationships
• Separate deterministic work (validation, parsing, math) from flexible interpretation
• Load knowledge progressively — only what's needed, when it's needed

The result: Claude works like a trained specialist who understands both the domain AND your specific systems.

Installation

Claude Desktop (Recommended)

The packaged .zip file is included in this repository for easy installation:

1. Download create-expert-skill-v2.2.zip from this repository
2. Open Claude Desktop → Settings → Capabilities
3. Under Skills, click Upload Skill
4. Drag and drop the .zip file (no need to unzip)

Claude Code

Skills can be installed at user or project level:

Personal skills (available in all projects):

# Unzip and copy to your personal skills directory
unzip create-expert-skill-v2.2.zip -d ~/.claude/skills/

Project skills (shared with team via git):

# Unzip into your project's .claude/skills/ directory
unzip create-expert-skill-v2.2.zip -d ./.claude/skills/
git add .claude/skills/create-expert-skill
git commit -m "Add create-expert-skill"

Claude Code automatically discovers skills in these locations.

Claude Agent SDK

For programmatic usage with the Agent SDK:

1. Create skills directory in your project: .claude/skills/
2. Unzip the skill into that directory
3. Enable skills in your configuration by adding "Skill" to allowed_tools

from claude_agent_sdk import query, ClaudeAgentOptions

options = ClaudeAgentOptions(
    allowed_tools=["Skill", "Read", "Write", "Bash"],
    # Skills are auto-discovered from .claude/skills/
)

See Agent Skills in the SDK for full documentation.

Usage

Start a conversation:

"I want to create a skill for validating LEDES billing files"

Claude guides you through:

Assess  → Is this worth creating? (3+ uses, consistent procedure)
Discover → What's the domain expertise? What are the system ontologies?
Design  → What needs scripts vs guidance vs reference material?
Create  → Generate the skill
Refine  → Iterate until complete
Ship    → Package for deployment

How It Works

Two Knowledge Streams

Production-ready skills require BOTH:

Domain Expertise — Industry knowledge that applies universally:

• Standards and their versions (e.g., LEDES 98B vs XML 2.0)
• Professional conventions and edge cases
• Validation rules from specifications

Ontology Understanding — System-specific structures:

• Company policies and constraints
• Entity relationships (timekeepers → IDs → rates)
• Data format variations unique to your systems

Progressive Disclosure Architecture

Skills load knowledge in layers, not all at once:

Layer 0 (~25 tokens)   → Description only, always visible
Layer 1 (~500 tokens)  → Core procedures in SKILL.md, loaded when triggered
Layer 2 (~1000+ tokens) → Deep reference in resources/, loaded selectively

Why this matters: A 2,000-token skill that loads everything wastes context. A layered skill loads 25 tokens until needed, then 500, then more only if required.

Deterministic Scripts

Anything that can be computed exactly should be:

Task	Without Script	With Script
Validate date format	LLM guesses (sometimes wrong)	python validate.py (always right)
Sum line items	LLM approximates	Script calculates exactly
Check against schema	LLM interprets	Script returns pass/fail

Scripts run at zero token cost — Claude executes them and uses the output.

Skill Structure

my-skill/
├── SKILL.md              # Layer 1: Core procedures (300-500 tokens)
├── scripts/              # Layer 0: Deterministic automation
│   └── validate.py
└── resources/            # Layer 2: Deep reference (loaded selectively)
    ├── schemas/
    └── ADVANCED.md

Token Optimization

Technique	Before	After	Savings
Scripts	500 tokens explaining logic	python scripts/validate.py	~450
Reference files	Inline schema (200 tokens)	Link to file	~185
Layer 2 split	Everything in SKILL.md	Split to resources/	~750

Packaging

This skill includes an automated zipping procedure In most cases, it runs on its own once the expert skill is finished, returning the plug-and-play .zip of the skill directly in conversation. If this doesn't run automatically, simply ask Claude to deliver the packaged skill.

Files

create-expert-skill/
├── SKILL.md                          # Main skill (Layer 1)
├── README.md                         # This file
├── LICENSE                           # MIT
├── create-expert-skill-v2.2.zip      # Ready-to-install package
├── scripts/
│   ├── package_skill.py              # Packaging automation
│   └── README.md
└── resources/
    ├── templates/
    │   └── TEMPLATES.md              # Skill templates (minimal/enhanced/script)
    └── examples/
        └── EXAMPLES.md               # Domain patterns (billing, API, schemas)

Contributing

Found a bug or want to improve the skill?

• Open an issue for bugs or feature requests
• PRs welcome for templates, examples, or documentation

License

MIT — use freely, modify as needed.

Author

Vlad-Alexandru Nicolescu

---

Version: 2.2
Tested with: Claude Desktop