Optimizes LLM context windows and token usage through strategic attention-aware design and dynamic budgeting.
Context Engineering provides a specialized framework for managing large context windows in agentic systems, moving beyond simple prompt engineering. It addresses the 'Lost in the Middle' phenomenon by strategically positioning critical information where models pay the most attention. The skill organizes context into five distinct layers—System, Capability, Knowledge, Memory, and Observation—and implements automated token budgeting, sliding window compression, and progressive disclosure. This ensures high-signal performance while reducing token costs and preventing model degradation in long-running tasks.
主要功能
01Five-layer architecture for structured system, tool, and document management
02Compression triggers and sliding window memory management for history
03Automated MCP tool deferral to save up to 7,200 tokens per session
0469 GitHub stars
05Attention-aware positioning to mitigate information loss in long contexts
06Dynamic token budget calculation for major models like Claude, GPT, and Gemini
使用场景
01Designing high-performance system prompts for multi-agent architectures
02Managing long-running agent conversations without exceeding context limits
03Optimizing RAG retrieval pipelines to load only high-signal documentation
