What problem does it solve? LLM agents degrade unpredictably as context grows: critical instructions get lost in the middle of long prompts, tool outputs balloon to dominate the window, and nominal token limits mask a much lower effective capacity. This Skill provides the foundational mental models, thresholds, and techniques needed to design, debug, and optimize context for agent systems. ## Core Features & Use Cases - Context anatomy guidance: Covers system prompts, tool definitions, retrieved documents, message history, and tool outputs, with concrete thresholds such as the 60-70% effective-capacity rule and the lost-in-the-middle effect. - Progressive disclosure and budgeting patterns: Explains three-level progressive disclosure, compaction triggers at 70-80% utilization, and sub-agent compression ratios of 1,000-2,000 tokens. - Executable utilities: Includes a Python module for token estimation, priority-aware context assembly, message truncation, structure validation, and lazy file loading. - Use Case: When an agent's reasoning quality collapses after 20-30 tool calls, use this Skill to diagnose that message history is consuming 70-80% of the window and implement observation masking plus compaction triggers. ## Quick Start Ask the agent to explain why my agent's performance degrades in long conversations and how to fix it using context budgeting and progressive disclosure.
