Lucek presents skills as portable units of functionality for agents such as Claude Code, Codex, Cursor and open-source terminal-based harnesses. Instead of repeatedly placing domain knowledge in prompts, a skill stores a focused workflow in a directory whose primary Markdown file tells the agent what the skill does and how to use its supporting files.
The comparison with adjacent standards is explicit. Function calling lets a model select a defined tool and produce its arguments. MCP standardizes how tools are exposed and executed. Skills sit above those mechanisms by bundling the instructions, resources and scripts required for a particular behavior, while remaining able to invoke existing tools or MCP servers.
A skill needs a directory and a SKILL.md file with concise name and description metadata. The talk recommends progressive disclosure, short primary files, meaningful metadata, consistent terminology, examples and separate reference files for details. This keeps recurring context small while allowing an agent to load deeper material only when the task requires it.
The Apple Notes demonstration packages create, read, list and delete scripts with formatting guidance and documented limitations. The example shows both the practical value and the design work involved: authors must understand the workflow, define safe steps, record constraints and make the result repeatable before an agent can execute it reliably.
Watch on YouTube
