Architecture

Context Window

The token limit within which Claude operates per conversation, determining how much knowledge, history, and instructions fit in a single session.

The context window is the hard constraint that shapes every Knowledge OS design decision. Claude models have a fixed token budget (200K for most models) shared between instructions, conversation history, file contents, and output. A naive approach loads everything at once and hits the ceiling within minutes. Knowledge OS solves this with conditional loading (rule files load only when paths match), context budgets per PRD phase (typically 8-12 files max), and context compaction when the window fills. Understanding this limit is why skills load references selectively rather than dumping entire knowledge bases.

Where it shows up:

Session planningSkill designPRD phase budgetsArchitecture decisions

Build your Knowledge OS

90 minutes from zero to your first skill chain.