Agent Skills in the OpenAI API allow users to upload and manage reusable bundles of files, known as skills, which enhance the functionality of hosted or local shell environments. Each skill is structured as a folder containing a required SKILL.md manifest file that outlines its name, description, and other essential details. When you attach a skill to an execution environment, the API unzips the files, reads the SKILL.md, and integrates the skill's information into the model's context. This setup enables the model to execute specific scripts or instructions only when needed, keeping system prompts concise. Skills shine in scenarios where workflows require repeatable processes, particularly those that involve complex branching or conditional logic. They are well-suited for tasks like generating reports or performing data cleaning, where having a versioned and reusable set of behaviors is beneficial. Skills also support the inclusion of scripts and local assets, making them ideal for environments that need to execute code or handle specific formats without overloading the system prompts with detailed procedures. However, skills are not the right choice for one-off tasks or situations where real-time data is required. If a procedure is frequently changing, the benefit of using skills diminishes as they are designed for stable workflows. The article also clarifies the distinctions between skills, tools, and system prompts. System prompts define the overall behavior and constraints, while tools are for executing external commands or fetching live data. Skills are focused on executing repeatable workflows within a controlled environment. To create and use skills, you can upload them via the API using either a zip file or individual files. Each skill returns identifiers and version pointers, allowing for effective management and tracking of different versions. The article provides practical examples, like the CSV Insights Skill, which summarizes CSV data and generates reports and visualizations, alongside specific code snippets for implementation.