context-fundamentals_skill

Overview

This skill teaches foundational context engineering for agent systems, covering context components, attention mechanics, progressive disclosure, and context budgeting. It equips engineers to design, debug, and optimize how agents manage tokens and external knowledge. Use it as a primer before designing multi-agent architectures or production agent workflows.

How this skill works

The skill breaks context into core components: system prompts, tool definitions, retrieved documents, message history, and tool outputs, then explains how each impacts attention and token budgets. It explains attention-budget constraints, position-encoding effects, and the progressive disclosure pattern for loading information just in time. Practical guidance includes filesystem-based access, hybrid loading strategies, and explicit context budgeting and compaction triggers.

When to use it

• Designing or iterating on an agent architecture that must manage limited context
• Debugging unexpected agent behavior that may stem from poor context curation
• Optimizing token costs and model performance for long-running tasks
• Onboarding team members to context engineering principles and trade-offs
• Deciding where to place critical information in prompts and outputs

Best practices

• Treat context as a finite resource; favor high-signal tokens over exhaustive dumps
• Use progressive disclosure: load summaries first, full content only when needed
• Organize system prompts into clear sections (background, instructions, tools, outputs)
• Place critical data at attention-favored positions (beginning or end of context)
• Implement compaction and retention policies; trigger them at ~70–80% utilization
• Measure attention usage and assume context degradation as length grows

Example use cases

• Designing system prompts and tool definitions for a code-assistant agent
• Implementing just-in-time document retrieval for a customer support agent
• Creating compaction rules for long-running planning or debugging sessions
• Setting context budgets and monitoring token usage in production
• Choosing hybrid pre-load vs. on-demand strategies for static vs. dynamic data