What problem does it solve? Long conversations and verbose tool outputs quickly exhaust limited LLM context windows, degrading response quality, increasing latency, and raising token costs. This Skill provides systematic techniques to extend effective context capacity without requiring larger models. ## Core Features & Use Cases - Compaction Strategies: Summarize old conversation turns and tool outputs when approaching context limits, preserving key decisions and findings while cutting token usage by 50-70%. - Observation Masking: Replace verbose tool outputs with compact references once they have served their purpose, reclaiming the 80%+ of tokens that tool outputs often consume. - KV-Cache Optimization: Order prompts with stable content first (system prompts, tool definitions) to maximize prefix cache hits and reduce latency and cost. - Context Partitioning: Split work across sub-agents with isolated contexts so each subtask operates in a clean window. - Use Case: A long-running coding agent hits 80% context utilization mid-task. Apply compaction to summarize early turns, mask stale tool outputs, and reorder the prompt for cache stability, letting the agent continue without quality loss. ## Quick Start Ask the agent to audit the current conversation's token usage and apply compaction and observation masking to reduce context consumption below 70% of the limit.
