I Gave Claude New Skills - Teach, Don’t Prompt

Key Concepts

• Agent Skills (Cloud Skills): A feature allowing agents to acquire and utilize new knowledge, similar to how humans learn documentation for new tools.
• Model Context Protocol (MCP): A method for providing information to models, contrasted with agent skills which progressively load relevant information.
• Agent Development Kit (ADK): A multi-agent framework used in the example to build a customer support system.
• Clerk: A platform for user authentication and management, used to enable login for customer support agents.
• NexJS: A React framework used for building the front-end of the application.
• Skill Creator Skill: A built-in skill within Claude that can generate new skills by processing provided documentation.
• YOLO Mode: A mode of operation for Claude code, implying a more direct or experimental approach.
• Parallel and Sequential Agents: Different types of agents that can operate simultaneously or in a specific order within a multi-agent system.
• Gemini SDK: Software Development Kit for Google's Gemini models.
• Gemini 2.5 Flash: A specific version of the Gemini model.
• Superbase: A platform that can integrate with Clerk for backend services.

Building Agent Skills for Cloud Code

The video demonstrates a practical approach to building and integrating agent skills for cloud code, enabling it to handle unseen packages and coding frameworks. The speaker posits that "agent skills or cloud skills" are more important than "model context protocol."

Full Stack Application Implementation Overview

The project involves building a full-stack application with:

• Front-end: NexJS
• Authentication: Clerk, allowing Gmail login.
• Back-end Logic: Utilizing the Agent Development Kit (ADK) through a custom skill provided to Claude.

This method differs from approaches like sub-agent or MCP by progressively loading necessary information rather than dumping all tools or context into the context window, mirroring how humans learn new documentation.

Implementing a Custom Skill for ADK

1. Skill Creation Process:

• The ADK documentation was downloaded and zipped.
• This zip file was provided to Claude's "skill creator skill" with a prompt to create a skill for ADK.
• Claude analyzed the HTML files within the zip, understood the structure, and used its skill creator to generate a new skill.
• The output is a zip file containing the new skill and explanatory files.

2. Skill Structure:

• skill.md: Contains a quick start guide for Python and Java, and core concepts.
• Reference Files: Built upon the provided documentation.
• Example Files: Demonstrate various ADK features.

3. Local Integration:

• The created skill zip file was downloaded.
• A skills folder was created within a claude folder in the project.
• The extracted skill folder was placed inside the skills folder.
• The skill.md file within the skill folder provides initial descriptions, quick start guides, and references.

Building a Multi-Agent System with ADK

1. Project Goal:

• To build a multi-agent system for customer support using ADK.
• To integrate Clerk for authentication, allowing each customer support representative to log in and use their own set of agents.

2. Initial Prompt to Claude:

• "Help me create a multi-agent system for customer support. This is supposed to be a customer support system for an e-commerce website. Create dummy data sets that we can interact with. Customers should be able to look up different product prices for those products. They should be also able to return products that they have purchased and also inquire about state of the status of their products or orders. Look at the documentation of the new ADK skill that I have provided and based on that create a multi-agent framework for us."

3. Refinements and Iterations:

• The initial prompt was refined to include "use both parallel and sequential agents."
• The speaker noted the intention to use the ADK UI, with a future plan for a custom UI for agents.
• Claude initially encountered issues with API configurations but was able to correct them through back-and-forth interaction.
• The model was updated to "Gemini 2.5 flash."
• The implemented system had 14 different tools available to the agent.

4. Dummy Data and Agent Functionality:

• Dummy data was created for customers and orders.
• The system implemented a single customer support agent with multiple tools, fulfilling the initial requirements.

5. Demo of Agent Capabilities:

• Product Lookup: The agent could search for available headphones and report stock quantities.
• Customer Order Inquiry: The agent could find a customer named Michael, identify their order, and state the total amount.
• Complex Product Inquiry: The agent could search for SSDs, their pricing, and check if any customers had ordered them. The speaker noted a need for more tools to search orders by items, not just customer IDs.

Integrating Clerk for Authentication

1. Purpose:

• To add authentication to the application, allowing customer support representatives to log in with their accounts.
• Clerk is highlighted as an excellent platform for authentication and user management, with a generous free tier.

2. Clerk Setup Process:

• Create an account on Clerk.
• Create a new application within the Clerk dashboard.
• Configure sign-in options (e.g., email, Google).
• Name the application (e.g., "agent v2").
• Copy the API keys from Clerk and add them to the project's .env file.
• Clerk provides a copyable prompt with integration instructions for Next.js apps.

3. Integration with Claude:

• The Clerk integration instructions were pasted into Claude code.
• Claude was prompted to integrate this into the existing Next.js app.
• There were some back-and-forth interactions to modify the API implementation for ADK, but the integration was successful.

Final Implementation Demo

1. User Interface:

• The final implementation presents an interface for human customer support representatives.
• The AI, powered by Google ADK, assists them in retrieving information.
• Authentication is handled by Clerk.

2. Login and Agent View:

• Users can log in using Google or a custom email.
• Upon login, the customer support agent sees an interface with analytics and an AI chat assistant.
• The system is powered by Google ADK.

3. Clerk Dashboard:

• The Clerk dashboard provides insights into user management and application details.

Conclusion and Future Potential

The process of using agent skills to provide new documentation to Claude for feature implementation is described as "tedious" but "extremely promising." This approach is expected to be highly beneficial for implementing company-specific Standard Operating Procedures (SOPs), such as those for code reviews. The speaker also recommends a previous video on agent skills for further learning.