Skills have emerged as a key feature in Claude Code, providing users with flexibility and ease of creation. However, this very flexibility can lead to confusion regarding their effectiveness. Questions often arise about which skills to develop and how to create high-quality ones. Anthropic has been utilizing skills extensively within Claude Code, with hundreds currently in operation. The following insights reflect our experiences.
Skills are not merely markdown files; they are intricate folders that can contain scripts, assets, and data. An agent can discover, explore, and utilize these resources. Within Claude Code, skills come with a variety of configurable options, including dynamic hook registrations. The most effective skills leverage these features and folder structures in innovative ways.
Through our analysis, we have identified several recurring categories for skills. High-performing skills typically fit neatly into one of these categories, while those that blur the lines often lead to confusion. Although this list is not exhaustive, it serves as a useful guide for identifying potential gaps in your organization.
1. Library and API References: Skills that clarify the correct usage of libraries, CLI, or SDKs, encompassing both internal and public libraries. These often include folders with code examples and a list of common pitfalls.
2. Product Verification: Skills that guide the testing and verification of code functionality, often in conjunction with external tools. These are crucial for maintaining accuracy and may justify dedicating an engineer to refine them.
3. Data Acquisition and Analysis: Skills that connect to data and monitoring stacks, containing libraries for data retrieval along with dashboard IDs and workflow instructions.
4. Business Processes and Team Automation: Skills that automate repetitive tasks, with the ability to save past execution results in log files for context continuity.
5. Code Templates and Scaffolding: Skills that generate boilerplate code tailored to specific functions in the codebase, often combined with composable scripts.
6. Code Quality and Review: Skills that ensure code quality within the organization and assist with reviews, potentially including deterministic scripts for reliability.
7. CI/CD and Deployment: Skills focused on fetching, sending, and deploying code, which may reference other skills for data gathering.
8. Runbooks: Skills that take a symptom (like a Slack thread or alert) and conduct an investigation using various tools to produce a structured report.
9. Infrastructure Operations: Skills designed for routine maintenance tasks and operational procedures, some of which involve destructive actions requiring safeguards.
To create effective skills, we recommend avoiding obvious information since Claude Code already understands much about your codebase. Focus on insights that push Claude beyond conventional thinking. Additionally, incorporating a "gotchas" section that outlines common issues encountered can significantly enhance the value of any skill.
Recently, we also launched Skill Creator, a tool designed to simplify the process of creating skills within Claude Code.
The introduction of this comprehensive guide is set to enhance the market for AI development tools, potentially raising the bar for competitors as companies strive to optimize their own skill creation processes.
Informational material. 18+.
