Anthropic published a new feature called Skills and on the surface it sounds like a simple plugin of some sort. What it actually does is JIT context injection with increasing levels of information added as needed:
It uses YAML frontmatter for token efficiency and has the following anatomy:
Level 1: Metadata (always loaded) < 100 Tokens
Level 2: Instructions (loaded when triggered) < 5000 Tokens
Level 3: Resources and code (loaded as needed) (∞)!
Read their full engineering blog post:
Equipping agents for the real world with Skills
Discover how Anthropic builds AI agents with practical capabilities through modular skills, enabling them to handle complex real-world tasks more effectively and reliably.
Some people will consider this a proprietary approach but I have used heavily to allow for larger protocols but seeing Anthropic standardise is great validation of the concept. You can find an excerpt of my AGENTS.md that uses this approach here
Very excited to see more of this "Context Engineering" in the coming months!