Claude Code Operating System

Most people start by putting everything into one giant instruction file. That works until the file becomes a junk drawer — important guardrails compete with API trivia, old notes stay in the prompt forever, and the model sorts through irrelevant context before doing anything. Here the main contract stays focused on two jobs: the judgment rules that affect nearly every task, and a routing table that points to the file that owns each area's knowledge. Memory works the same way — an index, not the memory itself, one line per lesson linking to a focused file with the evidence and how to apply it.

Rules, context, procedures, and memory stay separate. Always-on rules control judgment and stay short, because they spend context on every task. Context files explain the business and load only when the task matches. Tool docs record observed reality — live behavior beats remembered documentation. References and commands encode procedures, and memory holds durable lessons with edits gated by approval.

The operational conventions are what make an agent reliable: checkpoint everything, so a crashed run resumes instead of re-spending credits; read before write, so the agent never clobbers a manual edit; knowledge lives in one place, so updating it once means every workflow inherits it; and the human gates are load-bearing, not decorative.

The useful part is not that the assistant remembers more. It's that failures become infrastructure: when an API behaves differently than documented, the immediate task gets fixed and the reference gets corrected; when a workflow reveals a missing check, the check goes into the shared procedure; the next project inherits the correction automatically. The model can change — the durable advantage is the context architecture around it.