Context Fundamentals

This is the foundational reference for understanding how language models actually use context: system prompts, tool definitions, retrieved documents, message history, and tool outputs. The core insight is treating context as a finite attention budget rather than just a token limit. It explains why throwing more context at a problem often makes things worse, covers the progressive disclosure pattern for loading information only when needed, and walks through practical tradeoffs like tool outputs consuming 80%+ of tokens in typical agent runs. Use this when designing agent architectures, debugging weird behavior, or explaining to your team why that 200k context window won't magically solve your memory problems. The section on prompt altitude is especially good.