Context Engineering Skill

Purpose

Establishes a production-grade context management framework for agent execution loops. Prevents prompt token overflows, minimizes runtime API costs, maintains execution states across restarts, and mitigates semantic drift during long chat sessions. This system provides automated mechanisms to prune, summarize, score, and inject context blocks dynamically into prompt structures.

Core Principles

1. Token Allocation as Budget Management: Treat model context windows as finite memory caches. Allocate tokens dynamically based on priority weightings.
2. Crash-Resilient State Persistence: Store execution progress parameters externally. If a process errors or restarts, it must be able to restore context and resume.
3. Semantic Anchoring: Maintain a constant semantic anchor to the initial goal text. Actively prune conversational drift.
4. On-Demand Context Loading: Do not load documents statically. Query vector databases or status flags dynamically based on the current state.
5. No Token Waste: Strip comments, reduce whitespace, and format structural tables compactly.

Agent Protocol