Claude Skills & Agentic Patterns

This repository contains the example code for my Medium article https://medium.com/@krish777/claude-skills-the-hidden-superpower-behind-smarter-ai-agents-b40618cc1bbe on building with AI Skills. It demonstrates three distinct patterns for equipping LLMs with specialized knowledge and tools.

Key Examples

1. manage_skills.py (Cloud Native)

Demonstrates Anthropic's Native Skills Beta. This pattern uploads skill folders to the cloud, giving Claude "native" access to instructions and scripts via a unique ID.
- Best for: Speed and production-ready Claude-exclusive tools.

2. unit_converter_agent.py (Dynamic Loading)

Implements the Discovery → Loading pattern from agentskills.io. It shows the model a metadata "menu" and only injects full instructions when the model requests them.
- Best for: Token efficiency and massive skill libraries.

3. universal_agent.py (Model Agnostic)

This script takes the dynamic loading pattern and makes it compatible with GPT-4, Gemini, and Claude with just a single line change.
- Best for: Multi-model support and avoiding vendor lock-in.

Setup

1. Install Dependencies:
   This project uses uv for package management.
   bash uv pip install anthropic litellm python-dotenv

2. Environment Variables:
   Create a .env file based on .env.example and add your API keys:
   text ANTHROPIC_API_KEY=your_key_here OPENAI_API_KEY=your_key_here

3. Running Examples:
   bash python manage_skills.py python universal_agent.py

Skills Structure

Each skill lives in custom_skills/ and follows a standard format:
- SKILL.md: Instructions with YAML frontmatter.
- scripts/: Python/JS code the skill can execute.
- resources/: Data files (CSV/TXT) the skill can reference.