How to build AI skills instead of many agents

TL;DR

A skill lets one AI agent operate a growing toolbox instead of spawning a new agent for every task. A skill is a folder with skill.md that tells the agent when and how to use it. You can build skills manually, have Claude generate them, or record a workflow for Claude to learn.

Transcript

Anthropic's own engineers just said they stopped building a new agent for every single job. And that sounds completely backwards until you see what they switched to building instead. Think about your phone. The AI model is the processor. The AI agent running it is the operating system. And skills skills are the apps on your phone. You don't buy a w... Read More

Key Insights

• A skill is essentially a folder with a single file called skill.md, serving as a reusable tool for an AI agent.
• The skill.md file’s two header lines, name and description, trigger when the skill is used and how the agent should apply it.
• Progressive disclosure is the mechanism that keeps tools hidden until required, preserving context.
• You can create a skill manually by writing a skill.md and adding a description that instructs the agent how to execute the task, enabling repeatable performance.
• Alternatively, you can let Claude build the skill by answering a series of questions to define the workflow, then save the generated skill for reuse.
• A second method is to record your screen so Claude learns a workflow directly from your actions, which can capture complex steps precisely.
• A YouTube research skill can be built to fetch viral pockets and extract patterns from video titles using data sources like vidIQ.
• Skills functions as an app-like library for an AI agent, allowing a single agent to perform more tasks without creating separate agents for each job.

Questions & Answers

Q: What is a AI skill in this context?

An AI skill is an app-like component that teaches an agent how to perform a repeatable job by using a specific file, usually skill.md, that holds the description and instructions. It enables progressive disclosure so the agent only activates the skill when the user requests it, making workflows more scalable and efficient.

Q: How is a skill.md file structured in this approach?

The skill.md file starts with two header lines: the name of the skill and a description that guides when the skill should be used. Below those lines are plain English instructions detailing how to perform the job every time. This simple structure makes the skill easy to read and modify.

Q: What is progressive disclosure and why is it important?

Progressive disclosure is the idea that the toolbox stays closed until a specific skill is needed. This reduces context load on the AI by not reading every skill up front, allowing the agent to remain focused on the task at hand and only accessing the relevant skill when appropriate.

Q: What are the three ways to build a skill described in the video?

The video explains three methods: manually writing a skill.md with name, description, and instructions; asking Claude to interview you and generate the skill based on your answers; and recording your screen so Claude can learn a workflow directly from your actions.

Q: How can Claude assist in building a skill?

Claude can either generate the skill by asking targeted questions about the workflow or by using a starter prompt to guide the creation. You provide answers, and Claude returns a completed skill that you can save and reuse, streamlining the creation process.

Q: What is the benefit of using a skill instead of creating new agents for every job?

Using a skill consolidates repeatable tasks into a reusable module, reducing the need to spawn new agents for each job. This approach lets a single agent grow a toolbox of capabilities, improving efficiency and consistency while saving time on monitoring, training, and maintenance.

Q: What role does the YouTube pockets example illustrate?

The YouTube pockets example shows how a skill can be built to search YouTube niches, filter by recency and quality signals, identify viral formats, and extract top titles. This demonstrates turning a complex manual workflow into a repeatable, reusable skill that can ideate video ideas.

Q: Where can you find existing skills and what should you check before installing third-party ones?

Existing skills can be found in Claude libraries, GitHub, Claude Marketplaces, and community repositories. Before installing third-party skills that contain code or scripts, you should perform a security check to verify provenance and ensure there are no malicious components.

Summary & Key Takeaways

• A skill is a folder containing a single file that defines how an agent should perform a repeatable task.

• The critical file is skill.md, which holds the name, description, and instructions used by the agent.

• Progressive disclosure keeps the toolkit closed until a skill is explicitly needed, reducing context loading and improving efficiency.