optimizing-context_skill

Overview

This skill provides a unified interface for context engineering patterns used to optimize AI agent context. It consolidates compression, degradation detection, KV-cache optimization, and session management into a single, actionable entry point. Use it to keep agents performant, cost-effective, and robust over long or complex workflows.

How this skill works

When invoked the skill inspects the agent’s context window, usage thresholds, and recent tool outputs to classify the current problem (e.g., window saturation, degradation, cache inefficiency, session drift). It then selects and applies the appropriate technique—observation masking, summarization/compaction, degradation mitigation, KV-cache best practices, or session handoff—and emits a short metrics report. The skill supports progressive disclosure so operators see low-friction fixes first and aggressive measures only when needed.

When to use it

• Context window is approaching capacity or token budgets are high
• Agent starts ignoring or contradicting mid-context facts (lost-in-middle)
• You want to cut API costs by optimizing repeated context (KV-cache)
• Sessions must be persisted or handed off cleanly across runs
• Long-running or multi-phase tasks that risk context drift

Best practices

• Monitor utilization bands (<60%, 60–80%, 80–95%, >95%) and escalate compression progressively
• Prefer observation masking for large tool outputs and summarization/compaction for mixed content
• Keep system prompts stable, use append-only history, and deterministic serialization for KV-cache effectiveness
• Maintain a scratchpad and todos for critical decisions; update after major tool calls
• Use plan mode and recurring reminders to anchor objectives during long workflows

Example use cases

• Compressing verbose tool outputs into reference links to save tokens and preserve traceability
• Detecting 'lost-in-the-middle' behavior and reciting objectives via todo updates to restore focus
• Applying KV-cache rules (stable prefix, append-only) to lower per-request tokens and cost
• Persisting session state into a structured context directory for handoff between agents or runs
• Compacting older conversation history while retaining phase checkpoints for audits