Claude Skills: Glimpse of Continual Learning?

Key Concepts

• Skills: Custom onboarding material for Claude, packaging expertise to make it a specialist in specific areas. They are a set of instructions and resources for repeatable workflows.
• MCP Servers (Multi-tool Calling Protocol Servers): A way to connect tools to Claude, but they load all available tools upfront, consuming significant context.
• Sub-agents: Specialized agents that execute instructions, but their context is isolated from the main agent, and only the final output is passed back.
• Skill.md: A markdown file that defines a skill, containing the system prompt, capabilities, usage instructions, input/output formats, and examples.
• Progressive Disclosure: The core design principle of skills, where information is loaded only as needed, similar to a well-organized manual.
• Continual Learning (without model weight modification): Skills enable Claude to learn new SOPs and capabilities by adding new skills, without retraining the model.
• Composable: Skills can be automatically identified and coordinated by Claude, unlike static knowledge bases in custom instructions or projects.
• Slash Commands: Custom workflows that can be triggered manually.

Understanding Claude Skills

This video introduces Anthropic's new "skills" feature for Claude, which is presented as a potentially significant advancement, possibly surpassing MCP servers. Skills are designed to enable repeatable workflows where an agent must adhere to specific Standard Operating Procedures (SOPs).

What are Skills?

According to Anthropic, skills are akin to "custom onboarding material" that allows users to "package expertise," transforming Claude into a specialist in areas relevant to the user. This is described as an early form of continual learning without altering the model's weights.

Fundamentally, a skill comprises a set of instructions and resources that an agent can utilize to execute a defined, repeatable pattern. These are organized within a folder structure, typically including a skill.md file and other supporting files that the agent can reference when tackling specific problems.

Internal Use Case: File Creation

Anthropic uses skills internally, with an example being their file creation capability. This skill is defined by a skill.md file and associated Python scripts that the agent executes to create or modify files.

Skills vs. MCP Servers and Sub-agents

While MCP servers and sub-agents can also be used for similar purposes, skills offer distinct advantages, primarily in how they manage the context window.

MCP Servers and Context Management

When an MCP server is connected, it loads all its available tools, consuming a substantial portion of the context window (e.g., 32,000 tokens, approximately 16% of the context in the example). The agent then needs to determine which tool to use at each step.

Sub-agents and Context Isolation

Sub-agents are specialized but operate with isolated contexts. When a sub-agent is called, it performs its tasks, but only the final outcome is returned to the main agent. The intermediate operations and context are not passed on.

Claude Skills and Context Management

Claude skills differ significantly due to the skill.md file. This file contains the system prompt, defining the skill's purpose, and lists the tools available to that specific skill.

Example: Financial Ratio Analysis Skill

A custom skill provided by Anthropic for financial ratio analysis includes a skill.md file that describes its capabilities, usage, input/output formats, and examples. While this might appear similar to MCP servers, skills are specialized for repeatable workflows, and the agent is explicitly instructed on how to use the tools and in what order.

Structure of skill.md:

1. Description/Metadata: A brief overview of the skill's function (approx. 150 tokens).
2. Capabilities/Tools: Detailed information about available tools and how to use them (less than 5,000 tokens).
3. Actual Files/Resources: The core implementation files, which can potentially have unlimited token capacity.

Context Loading with Skills:

When an agent uses skills, it initially loads only the description or metadata (around 100-150 tokens) for each skill. Upon receiving a user request, the agent selects relevant skills based on similarity. If a skill is not relevant, it can be discarded. Only when a skill is deemed relevant does the agent load its body and begin using its tools.

Progressive Disclosure:

Anthropic highlights "progressive disclosure" as the core design principle of skills. This allows Claude to load information only as needed, making skills flexible and scalable, much like a manual that starts with a table of contents, then chapters, and finally an appendix. This is a significant benefit for context management.

Skill Components and Continual Learning

Skill Structure

A skill can include references to other markdown files, enabling a hierarchical structure for instructions. The skill.md file guides the agent on how to perform actions and utilize available tools.

Continual Learning Mechanism

This structure allows a single skill to define multiple workflows. It is particularly useful for teaching agents new SOPs or company-specific guidelines. By introducing new skills, the agent's capabilities can be expanded without retraining the model, enabling a form of "on-the-fly" learning. This is presented as a crude but effective method of continual learning.

Skills vs. Custom Instructions and Projects

While custom instructions and projects in Claude and ChatGPT offer similar functionalities, skills offer key differentiators:

• Composability: Claude automatically identifies and coordinates the use of skills, unlike static knowledge bases in projects.
• Reusability: Skills can be used across different Claude applications, Claude Code, and APIs, and are potentially shareable.

Availability and Skill Creation

Skills are currently available on the API, Claude Code, Claude web interface, and the desktop app. Users can access pre-built skills created by Anthropic (e.g., "Canvas Design Brand Guidelines") and create their own.

Skill Creator Tool

Claude itself can be used to create skills via a "skill creator" tool. The process involves defining the skill's purpose, and Claude generates the necessary skill.md file and associated workflows.

Example: Code Review Skill Creation

The video demonstrates asking Claude to create a skill for thorough code reviews. Claude uses its skill creator to generate a skill.md file with different workflows for the agent to follow during a code review. This is highly beneficial for enforcing company-specific templates and guidelines. The generated skill can then be integrated into Claude Code.

Future Outlook and Industry Adoption

The long-term adoption of skills by the industry remains to be seen. However, Anthropic's approach to skills is praised for its engineering elegance. The video notes a potential divide, with some coding agents adopting standards like agents.md while Claude Code might take a different path. The emergence of industry standards for agent skills is anticipated.

Conclusion

Claude skills represent a novel approach to enhancing agent capabilities by enabling specialized, repeatable workflows. Their strength lies in efficient context management through progressive disclosure and their potential to facilitate continual learning without model retraining. The composability and reusability of skills across different platforms offer significant advantages for developers and organizations looking to customize and scale AI agent functionalities.