How does observation masking save tokens?

It replaces large, verbose outputs from tools or shell commands with short reference IDs or elided summaries, keeping the context window clean while preserving key findings.

What is context compaction in AI agents?

Compaction is the process of summarizing previous conversational turns and tool outputs to free up token space while maintaining the essential logic and state needed for future tasks.

Can this skill help reduce my Claude API costs?

Yes, by optimizing token usage and improving KV-cache stability, the skill reduces the number of tokens processed in each request, leading to lower costs and faster response times.

When should I use context partitioning?

Use partitioning when a single task becomes too complex for one context window; it splits work across isolated agents, each focusing on a subtask without carrying unnecessary context.

Context Optimization

Name: Context Optimization
Author: ngxtm

byngxtm

•

生産性とワークフロー

Maximizes the effective capacity of AI context windows using advanced compaction, masking, and caching strategies.

The Context Optimization skill empowers Claude to handle complex, long-running tasks by strategically managing token usage through compression, observation masking, and partitioning. By distilling high-fidelity summaries of previous turns and eliding verbose tool outputs, it prevents performance degradation and context overflow while significantly reducing inference costs and latency. This skill is essential for developers building production-scale agentic systems where maintaining reasoning accuracy over extended sessions requires high-quality context management and efficient budget allocation.

主な機能

01Context partitioning to isolate subtasks across multiple agent instances

02Strategic context compaction for high-fidelity summarization of history

03Observation masking to replace verbose tool outputs with compact references

04KV-cache optimization patterns to maximize prefix matching and reduce latency

05Budget-driven triggers for automatic context cleanup based on token usage

065 GitHub stars

ユースケース

01Managing token-heavy tool outputs and massive documentation retrieval tasks

02Reducing API costs and latency in production-grade multi-agent systems

03Scaling long-running coding agents that process large codebases and complex logic

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add ngxtm/devkit context-optimization

For use in Claude.ai and ChatGPT

Download Skill