#162 - Skills do Claude

Key Points

• A skill is a reusable step-by-step guide (like a recipe or manual) for an AI model to perform a specific task, stored as a markdown file (skill.md) with a header (name and description) and a body (manual of steps).
• The header helps the model choose which skill to use; the manual contains instructions, reference files, templates, policies, examples, scripts, or assets the model should follow.
• Skills allow non-programmers to create complex automations by combining MCPs (interfaces between systems) and step-by-step logic, replacing what would previously require a full software system.
• Skills should be created, read, and edited by the AI model itself, not by humans, because the model knows the best format and can avoid breaking logic; humans should ask the model to summarize or modify skills.
• Information storage formats are shifting from human-friendly (e.g., slides) to model-friendly (e.g., structured text with semantic sections) to improve AI comprehension and efficiency.

Summary

In this podcast episode, Felipe explains the concept of "skills" in the context of artificial intelligence, specifically for models like Claude, GPT, and Gemini. A skill is essentially a reusable manual or recipe that tells the AI how to perform a specific task, eliminating the need to rewrite instructions every time. He compares it to a restaurant chef who would have to rewrite the same recipe daily for a cook—a skill encapsulates that knowledge permanently.

The technical structure of a skill is simple: it is a markdown file (skill.md) with two main parts. The first part is the header, which contains the skill's name and a brief description of what it does. This header is used by the AI to navigate and decide which skill to apply based on the user's request. The second part is the manual, which includes the actual step-by-step instructions, reference files, templates, policies, examples, scripts, or any other resources the model needs to execute the task. Felipe emphasizes that skills are not magical; they are just structured step-by-step guides.

A key advantage of skills is that they empower non-programmers to create complex automations. Felipe gives the example of his girlfriend, who uses MCPs (Model Context Protocols) to integrate tools like Granola (for meeting recordings), Slack (for conversations), and Notion (for aggregating information). She created a skill that automatically gathers her daily meetings and Slack messages, summarizes them, and generates actionable items—all without writing any traditional code. Similarly, Felipe himself used a skill to analyze the performance of an internal agent system by first writing a Python script to store data, then creating a skill to generate summaries and track evolution over time. He also mentions a friend named Marra who built a skill to automatically improve prompts by evaluating outputs against a dataset in a loop.

Felipe stresses that skills are best managed through the AI itself. Humans should not write, read, or edit the skill.md files directly because the AI knows the optimal format and can avoid logical inconsistencies. Instead, users should ask the model to create, summarize, or modify skills. This reflects a broader shift from human-friendly information formats (like slides) to model-friendly formats (like structured text with semantic sections), which allows AI to process information more efficiently. He notes that while skills can replace simple software systems, they are not always cost-effective for deterministic tasks—a Python script would be cheaper than using an LLM for such cases. However, for rapid prototyping or infrequent tasks, skills are invaluable.

Finally, Felipe promotes his course, which includes modules on skills, Claude Code, and Claude Code Orque, and offers a discount code for listeners. He encourages the audience to use skills to automate repetitive tasks and unlock productivity, concluding with a reminder that the sky is the limit for what can be achieved with this approach.