context-optimization

This skill provides a systematic framework for context engineering, focusing on extending the effective capacity of limited LLM context windows. It is designed for developers and AI engineers building production-grade agentic systems who face challenges with context limits, rising token costs, and latency in long-running conversational or analytical workflows. By treating the context window as a finite resource, this skill enforces discipline in data curation to maintain model performance and coherence.

• KV-cache optimization: Strategies for ordering prompts and stabilizing prefixes to maximize inference engine cache hits, significantly reducing latency and compute costs.

• Observation masking: Techniques to replace verbose tool outputs with compact references once their utility is served, preserving critical information while reclaiming significant token budget.

• Context compaction: Intelligent summarization of conversation history and tool outputs when utilization thresholds are met, prioritizing high-signal data while anchoring model behavior with stable system prompts.

• Context partitioning: Methodologies for decomposing complex tasks into smaller, isolated sub-agent contexts to avoid window saturation and improve reasoning quality.

• Activate this skill when encountering context-related constraints, such as token limits, degradation in reasoning accuracy for long trajectories, or the need for cost reduction at scale.

• Expected inputs include raw prompt data, conversation trajectories, and tool-use records; outputs are optimized, refined versions of these inputs that maintain task-critical state.

• Practical constraints: Compaction is a lossy operation; always measure effectiveness before and after applying strategies. Prioritize KV-cache stabilization as it is zero-risk, whereas aggressive compaction should be applied carefully to avoid loss of context-sensitive information. The skill is platform-agnostic, applicable to environments like Claude Code or standard LLM SDKs.