context-fundamentals

Context engineering is the discipline of curating the smallest high-signal token set that maximizes the likelihood of desired model outputs. It treats context as a finite attention budget covering system prompts, tool definitions, retrieved documents, message history, and tool outputs.

The U-shaped attention curve means models recall information at the beginning and end of context more reliably than the middle, with middle positions showing 10-40% lower recall accuracy. Place critical constraints at the top or bottom of prompts, never in the middle.

No. Effective capacity is typically well below the nominal window because attention quality degrades as a gradient, not a cliff. Processing cost also grows disproportionately with length, so smaller high-signal context outperforms larger low-signal context.

The 4 characters per token heuristic works only for English prose. Code tokenizes at 2-3 characters per token, URLs and paths tokenize per punctuation mark, and non-English text often runs 1-2 characters per token. Use the provider's actual tokenizer for budget-critical calculations.

Trigger compaction at 70-80% context utilization, before message history crowds out active instructions. In agentic loops, history can silently consume 70-80% of the window after 20-30 tool calls while reasoning quality collapses without visible symptoms.

Progressive disclosure loads only skill names and summaries at startup, fetching full content only when a task explicitly matches activation conditions. It applies at three levels: skill selection, document loading, and tool result retention, preventing context stuffing from eager loading.