• Fully compatible with existing skills: existing skills can be copied directly, no change needed
• Framework-free: can be used without any framework, or with other frameworks (currently only compatible with LangChain - more coming in the future!)
• Model-agnostic: Works with any LLM
• Multi-source skill discovery: From project, Anthropic config, plugins, and custom directories with priority-based conflict resolution
• YAML frontmatter parsing with comprehensive validation
• Progressive disclosure pattern (metadata-first loading, 80% memory reduction, Level 2 content caching)
• Script execution: Execute Python, Shell, JavaScript, Ruby, and Perl scripts from skills with comprehensive security controls
• Plugin ecosystem: Full support for Anthropic's MCPB plugin manifests with namespaced skill access
• Nested directory structures: Discover skills in any directory hierarchy up to 5 levels deep
• Security features: Input validation, size limits, suspicious pattern detection, path security, secure file resolution, script sandboxing

---

What Skills Are

Agent Skills are modular capability packages that work like "onboarding guides" for AI. Each skill is a folder containing a SKILL.md file (with YAML metadata + Markdown instructions) plus optional supporting files like scripts, templates, and documentation. The Agent autonomously discovers and loads skills based on task relevance using a progressive disclosure model—first reading just the name/description metadata, then the full SKILL.md if needed, and finally any referenced files only when required.

Why Skills Matter

- Transform AI from assistant to operational team member — Skills let you encode your organization's procedural knowledge, workflows, and domain expertise into reusable capabilities that Claude can invoke autonomously. Instead of repeatedly prompting Claude with the same context, you create persistent "muscle memory" that integrates AI into real business processes, making it a specialized professional rather than a generic chatbot.

- Combine AI reasoning with deterministic code execution — Skills can bundle Python, Shell, JavaScript, Ruby, and Perl scripts alongside natural language instructions, letting AI use traditional programming for tasks where LLMs are wasteful or unreliable (like sorting lists, filling PDF forms, or data transformations). This hybrid approach delivers the reliability of code with the flexibility of AI reasoning, ensuring consistent, auditable results for mission-critical operations.

- Achieve scalable efficiency through progressive disclosure — Unlike traditional prompting where everything loads into context, skills use a three-tie