A plugin library for agents that forget the middle of everything

What it does

This repository is a set of fifteen platform-agnostic “skills”—structured guides and patterns—for building production AI agents that don’t collapse under their own context window. It treats the context window not as a simple token bucket but as a scarce attention budget, covering everything from compression strategies and multi-agent orchestration to evaluation harnesses and formal BDI cognitive modeling.

The interesting bit

The skills are packaged as plugins for Claude Code and Cursor, activating automatically when your task matches their domain; at startup, agents see only skill names and descriptions, pulling full content only when needed. That meta-design—progressive disclosure to save context—practices exactly what it preaches.

Key highlights

• Covers the full lifecycle: context fundamentals, degradation patterns, memory systems, tool design, latent briefing, and autonomous harness engineering.
• Cited in academic research from Peking University and a CMU/Yale/JHU survey as foundational work on static skill architecture.
• Includes concrete examples such as a multi-agent X-to-book pipeline and an LLM-as-judge suite with 19 passing tests.
• Skills use Python pseudocode and conceptual patterns rather than framework-specific dependencies, so you can transplant them into any agent stack.
• Also explores formal cognitive modeling, translating external RDF context into agent beliefs, desires, and intentions.

Caveats

• The README is truncated in the provided source, so some example details and implementation guides appear incomplete.
• These are conceptual guides and pseudocode, not a drop-in library; you will need to implement the patterns yourself.
• Several recently added skills (hosted agents, BDI mental states) are marked NEW, suggesting the collection is still expanding.

Verdict

Worth bookmarking if you are designing multi-agent systems or debugging why your LLM loses the plot halfway through a long session. Skip it if you need a pip-installable framework with ready-made handlers.