context-fundamentals

Context-fundamentals serves as the prerequisite skill for engineers and developers building production-grade AI agent systems. It establishes the technical discipline of context engineering, which treats the LLM context window as a finite attention budget rather than an infinite storage bin. By mastering this skill, users can design architectures that mitigate the 'lost-in-the-middle' phenomenon and U-shaped attention curve degradation common in long-context tasks.

The skill provides actionable patterns for system prompt organization, effective tool definition design, and message history compaction. It guides developers on how to apply progressive disclosure—loading information only when necessary—to preserve model focus and reduce token expenditure. This is essential for anyone responsible for debugging unexpected agent behaviors or optimizing performance in complex, multi-turn AI interactions.

• Principles for informativity over exhaustiveness to maximize signal-to-noise ratio in agent trajectories.

• Guidelines for position-aware placement of critical constraints to capitalize on high-recall areas (beginning and end of context).

• Strategies for tool definition refinement, including proper handling of parameter defaults, usage context, and serialized JSON bloat.

• Methods for cyclically refining conversation history to prevent context saturation.

• Techniques for observation masking, replacing verbose tool outputs with concise references once processed.

• Activate this skill when initiating new agent architectures or performing code reviews on existing prompt chains.

• Use it during incident response when agents exhibit drifting behavior or difficulty following instructions due to context overload.

• Input requirements include your current system prompt, tool definitions, or reasoning trace snippets.

• Outputs provide structural refactoring advice, attention budget analysis, and heuristic-driven prompt optimization.

• Note that effective capacity for most models is typically 60-70% of the advertised window; plan accordingly to avoid performance cliffs.