context-optimization

Context Optimization Techniques

Apply tactical techniques to maximize useful information per token within the context window. Context optimization operates at the token level — it is the operational counterpart to the conceptual foundation in context-fundamentals and the failure diagnosis in context-degradation. Every technique trades implementation complexity for token savings.

When to Activate

Activate this skill when:

• Reducing token usage in production agent systems
• Implementing prefix caching for repeated context patterns
• Designing observation masking for tool outputs
• Partitioning context between static and dynamic sections
• Allocating token budgets across context components
• Optimizing retrieval precision to minimize irrelevant content

Do not activate this skill for adjacent work owned by other skills:

• Explaining why token efficiency matters: context-fundamentals.
• Diagnosing specific failure patterns: context-degradation.
• Designing compression strategies for long sessions: context-compression.
• Offloading content to filesystem: filesystem-context.
• Page load and Core Web Vitals regression: GStack /benchmark — that measures performance, not token savings (see this repo's benchmarks for context metrics).

Core Concepts

Partition context into static and dynamic regions. Static content (system prompt, tool definitions, persistent instructions) benefits from prefix caching — it is computed once and reused across turns. Dynamic content (user messages, tool outputs, retrieved documents) changes every turn and cannot be cached.

Apply observation masking to tool outputs: after processing, replace verbose outputs with compact summaries. The full output served its purpose during the processing turn; subsequent turns need only the conclusions.

Budget tokens explicitly across components. A 128K context window might allocate: 10K for system prompt, 5K for tool schemas, 20K for retrieved documents, 50K for conversation history, and 43K headroom. Monitor actual usage against budget and trigger compaction when any component exceeds its allocation.