I’m designing a personal AI OS for my Mac and would value a reality check from people who have built one beyond the demo stage. The goal is not another dashboard or a giant vault. It is a durable, human-readable operating layer that gives Codex/Claude the right context for each task while I retain ownership of the files. Current structure: AI OS/ - AGENTS.md — instructions for AI agents - AIOS.md — human-readable system map - context/ — personal context, business context, strategy, current priorities - projects/ — active outcomes with a clear end state - areas/ — ongoing responsibilities such as health, administration and communication - knowledge/ — long-term notes and discoveries - plans/ — dated implementation plans - outputs/ — generated deliverables - reference/ — reusable source material - source-maps/ — pointers to Git repos, email, cloud folders and sensitive sources - scripts/ — maintenance and automation - archive/ — inactive material Each serious project would contain PROJECT.md, NOW.md, HANDOVER.md, DECISIONS.md, SOURCES.md and local agent instructions. Important boundaries: - Git repos can remain independent rather than being forced into one monorepo. - The AI tool’s own runtime/config folder stays outside the AI OS. - Secrets never enter Markdown or Git. - Medical, legal and financial source files are not loaded by default; the OS contains indexes or source maps and loads raw material selectively. - Fresh sessions receive only the context required for the task, rather than the entire vault. My questions: 1. Is this the right separation between context, projects, areas, knowledge and source maps? 2. What files or maintenance habits turned out to be essential in your real AI OS? 3. Where does this approach become too complex or break down at scale? 4. How do you handle selective context loading and sensitive personal data in practice?