Skill UI

This skill details advanced techniques for enhancing the effective context capacity of LLMs. It outlines four core strategies—Compaction, Observation Masking, KV-Cache Optimization, and Context Partitioning—to prevent context limits from constraining task complexity. By strategically managing token usage, users can reduce operational costs, minimize latency, and implement robust, long-running agent systems for handling large documents or conversations.

This skill set provides advanced, structured strategies for optimizing Large Language Model (LLM) interactions. It covers techniques such as intelligent context summarization, dynamic trimming, context routing, and precise token budget allocation to prevent 'context rot'. It is essential for building robust, long-running, multi-turn conversational AI systems.

A comprehensive guide to optimizing context window usage within AI coding environments like Cursor. Learn how to strategically use @-mentions, manage context pills, and leverage different context sources (codebase, documentation, web) to prevent context overflow and ensure the AI generates highly accurate, relevant, and context-aware code suggestions and answers for complex development tasks.

A protocol that provides LLM agents with persistent structural memory of a codebase. It maps the project as a directed graph, tracking not only dependencies (what imports what) but also the 'reasons' (why it connects). This allows agents to understand context, perform impact analysis, and navigate large repositories without loading the entire source code into the context window.

This skill manages and refines search results obtained from Exa's neural search API. It provides comprehensive control over content extraction scope, allowing users to selectively retrieve metadata, highlights, full text, or structured summaries. Key features include implementing result caching with TTL, deduplicating citations, and crucial token budget management, ensuring the retrieved context fits within LLM context window limits for robust RAG pipelines.

This skill leverages Google Gemini's advanced image generation model to create high-quality, context-specific visuals. It is essential for developers needing placeholder images for UIs, generating detailed diagrams for technical documentation, or visualizing abstract concepts, logos, and assets across cross-platform projects (Windows, macOS, Linux).

Extends transformer context windows using RoPE, YaRN, ALiBi, and position interpolation so long documents (32k-128k+ tokens) can be processed, pre-trained models lengthened, and efficient positional encodings or extrapolation strategies deployed.

This skill provides methodologies for designing sophisticated, layered memory architectures for AI agents. It moves beyond simple context windows and basic vector stores by incorporating short-term, long-term, and graph-based memory systems. Implementing this is crucial for agents that require persistence across sessions, maintaining complex entity consistency, and performing advanced, relationship-driven reasoning over accumulated knowledge bases, including time-aware fact tracking.

This skill provides advanced strategies for managing large language model (LLM) context windows and token usage via OpenRouter. Functions include context window lookup, context-aware model selection, conversation trimming, document chunking for RAG, and prompt caching. Use this when dealing with long conversations or large documents to maximize quality while minimizing API costs and preventing context limit overflow.

Flash Attention optimizer delivers 2-4x speedup and 10-20x memory savings for transformer attention, ideal when training or running long-context models that hit GPU memory or latency limits; supports PyTorch native scaled_dot_product_attention, the flash-attn library, H100 FP8 precision, and sliding-window attention.

This skill provides advanced context management for AI in large-scale, complex codebases. It optimizes context windows, ensures memory persistence across multiple sessions, and helps prioritize critical information. Use it when working with massive code repositories to maintain context and improve developer workflow.

This pattern addresses the fundamental limitation of context window size by treating the filesystem as a persistent, external memory. Agents can write large outputs, plans, or findings to files, offloading them from the active context. Instead of carrying all information, agents use targeted search and retrieval tools to load only the necessary context on demand, drastically reducing token bloat and improving performance over long, complex tasks.