Video - "Skills, Rules, Subagents: Explained!"

Key takeaways

Two major areas of context: Static, based around text and prompts, and Dynamic, based around tool exposure
Everything was originally prompt-driven, and AI companies rallied to create Rules to prevent agents from hallucinating as much
MCP systems are how we expose third-party tools to agents, the issue is that with too many tools, you use up your context window
Since hooks are deterministic things you want to have happen, i.e., running scripts at certain points, we can log agent thoughts as they crank out code and how we can tweak our prompts, etc., to how they think.
- Agent analysis and recording with hooks!
Subagents are on the come-up, where you create a system prompt to define a specific agent type you'd like the "main" agent to spin up based on your requests to it; these subagents are given their own context windows separate from the main agent
Commands are prompts with persistence
Skills are souped-up commands where you can bundle commands and scripts within, and they're only called on demand, without taking up context window space
At the end of the day, Rules and Skills are the two foundational things to focus on
- Cursor is trying to push rules into more defined roles as project-level guard rails and exporting deterministic-ish things into Skills