The Only Claude Skills Guide You Need (Beginner to Expert)

Detailed Summary

Intro (0:00 - 1:18)

The video introduces Claude Agent Skills, a new feature from Anthropic, emphasizing the creator's prior experience in testing and understanding them. It outlines the video's scope, including explaining what skills are, their importance, how they differ from slash commands and MCPs, and how to build and optimize custom skills. The creator mentions that the video's title and thumbnail were generated by a personal assistant agent using a custom skill.

Deep Dive into What Skills Are (1:18 - 11:58)

• Current Challenge: Agents typically load all tool schemas (local or MCP) into the system prompt, leading to token inefficiency as only a fraction of tools might be used for a specific task.
• Skills as a Solution: Skills address this context engineering problem by introducing progressive disclosure.
• Skill Anatomy: A skill is essentially a folder containing a skill.md file, which includes YAML front matter (name, description) and detailed instructions. Optional files like scripts, markdown, text, or CSVs can provide additional context.
• Progressive Disclosure Explained: When a skill is given to an agent, only its name and description are initially loaded into the system prompt. The agent reads the full body of the skill.md file and other context only if it decides to use the skill, making it highly token-efficient.
• Skills vs. Slash Commands: Slash commands are user-triggered with predefined workflows, while skills are dynamically used by the agent, which also determines input arguments, offloading more work from the user.
• Skills vs. MCPs: While MCPs offer similar functionality (tools, resources, prompts), skills are simpler to implement (no server infrastructure needed) and leverage progressive disclosure more effectively, as MCPs typically load all tool definitions upfront.
• Key Benefits of Skills: Skills offer context engineering efficiency (handled by Claude harness), simplicity (markdown-based), and the ability to compound agent abilities by having skills leverage other existing skills.

Using Skills with Claude.ai (11:58 - 18:26)

• Accessing Skills: Skills can be managed in the Claude.ai web or desktop app by navigating to Settings > Capabilities > Skills.
• Default Skills: Claude comes with pre-built default skills from Anthropic, such as the 'skill creator' skill, which allows Claude to build new skills based on prompts.
• Uploading Custom Skills: Users can upload custom skills by zipping the skill folder (containing skill.md and other files) and dragging it into the upload section. The skill.md must have a valid YAML front matter.
• Demo: Cohort Analysis Skill: A custom skill is demonstrated to build a financial cohort model from transaction data. The agent examines the CSV, identifies the relevant skill, reads its instructions, and generates an Excel file with key metrics and formula-based calculations.
• Validation: The custom skill includes Python scripts to validate formulas in the generated Excel model, ensuring accuracy.

Using Skills with Claude Code (18:26 - 25:52)

• Integration: Skills are used similarly in Claude Code and the Claude Agent SDK by creating a skills folder within the agent's cloud directory, with each skill having its own subfolder.
• Portability: Skills are portable; the same skill folder can be zipped for Claude.ai or used directly in Claude Code.
• Anthropic's Skill Repo: Anthropic provides a GitHub repository with examples of their default skills (e.g., for PDF, Docs, PowerPoint, Excel) for reference.
• Demo: YouTube Thumbnail Skill Anatomy: A detailed walkthrough of a custom YouTube thumbnail creation skill's skill.md file.
• Detailed Description: The skill's description specifies when it should be used (e.g., creating/editing thumbnails).
• Organized Context: For complex skills, a table of contents in skill.md helps the agent quickly locate relevant context files (e.g., design requirements, prompting guidelines, optional assets).
• Mandatory vs. Optional Reading: Some context (like design requirements) is marked as mandatory for every use, while others (like thumbnail templates) are loaded only if the agent decides to use them, illustrating progressive disclosure in practice.

Demo of Custom Skill with Claude Code (25:52 - 34:13)

• Personal Assistant Context Engineering: The creator's personal assistant agent uses progressive disclosure by loading a top-level cloud.md file, which then points to subdirectories for memory, projects, and tools. Each subdirectory has its own cloud.md to provide more specific context as needed.
• YouTube Thumbnail Generation: The agent is prompted to generate title and thumbnail ideas for a YouTube video about Claude skills, referencing a research document with competitor video analysis.
• Agent Workflow: The agent uses MCP tools to gather data, curl commands to fetch existing thumbnails, and leverages its multimodal capabilities to understand image structures.
• Concept Generation: Claude analyzes the competition and suggests title and thumbnail concepts, applying design patterns from existing successful thumbnails.
• Thumbnail Creation: The agent uses the custom skill to generate thumbnails based on selected concepts. It invokes the Gemini CLI via bash commands, which is connected to the Nano Banana MCP server, to create the images.
• Review of Generated Thumbnails: The video shows the generated thumbnails, acknowledging that while some are good starting points, others require further refinement, demonstrating the iterative nature of AI-assisted design.

Optimizing Custom Skills (34:13 - 35:57)

• Iterative Improvement: Custom skills can be improved over time by iteratively prompting the agent. If a skill fails or doesn't meet expectations, users can provide feedback to Claude (e.g., "the skill failed because of X reason," showing expected vs. actual output).
• Agent-Assisted Debugging: Claude can then review the skill, identify failure points, suggest improvements, and implement fixes, allowing for continuous optimization.
• Future Content: The creator mentions an upcoming video that will deep dive into his personal assistant, which leverages these skills for everyday work.

Notable Quotes

"Conceptually, skills are actually really simple. It's just a folder." — Kenny

"This progressive disclosure makes the agent way more token efficient." — Kenny

"The main difference is that you have to trigger a slash command yourself, whereas with skills, they are used by the agent." — Kenny

"The second big difference is with progressive disclosure, which we just discussed. Claude is only going to see the name and description in the beginning, so the context load is minimal." — Kenny

"The nice thing is that the claude harness really takes care of how skills are used and the directions that go into the system prompt for leveraging skills." — Kenny

"I think that's really cool because you don't have to reinvent the wheel. you can just kind of piece different skills together and build new, more powerful skills that leverage other skills." — Kenny

"And that's the really cool thing about skills because claude code can just run terminal commands." — Kenny

"So, what's worked for me is telling cloud code or cursor or whatever you're using that the skill failed because of X reason and showing you it what the expected output was versus what happened and then having it review the full scale and identifying what it thinks the failure point was and then suggesting to you how to improve it and then going ahead and implementing that fix." — Kenny