Claude Cowork: Hierarchical CLAUDE.md Turns AI into Your OS
Build a persistent AI second brain using CLAUDE.md instruction files, memory.md for recall, and a 3-level folder hierarchy (root, workstations, projects) to automate email, finances, newsletters, and projects without burning rate limits.
CLAUDE.md and Memory.md Enable Persistent, Contextual AI Behavior
The core system relies on two plain-text Markdown files: CLAUDE.md as the instruction manual defining rules, and memory.md as a notepad for session-to-session recall. CLAUDE.md sets master rules like "at the start of every session, read memory.md before responding" and "when I say 'remember this,' write to memory.md." This creates persistent memory—tell Claude "current events distract from e-lists, remember that," and it adds an entry to memory.md's memory section, retrievable in future sessions via queries like "What did I say about distractions?"
A routing map table in root CLAUDE.md directs tasks to specific folders (e.g., email to Email HQ), while references point to resources only when needed, keeping token usage low. Voice principles.md (built by analyzing 30 Gmail emails or 5 writing samples) extracts patterns like "warm, direct, professional tone without stiffness," loaded before outputs for personalized content like newsletters matching your style. Active projects section in memory.md lists ongoing work (e.g., workshop outline, dinner plans) updated via commands, ensuring context across sessions.
Analogy: Root CLAUDE.md is the U.S. Constitution (applies everywhere); workstation CLAUDE.md files stack state laws on top for specialized rules. Limit root CLAUDE.md to 300 lines max, default to Sonnet model (1/5th Opus cost, sufficient 80% of time), and avoid rule duplication to minimize tokens.
3-Level Hierarchy: Root, Workstations, Projects for Scalable Specialization
Start with a root folder (e.g., "ClaudeOS") containing CLAUDE.md, memory.md, and 00-resources folder. Use Obsidian to view Markdown files readably (no learning curve needed). Download starter templates for these files.
Level 1 Workstations divide life areas: universal (e.g., Email HQ for cross-domain tasks) or dedicated (e.g., Personal Finances). Prompt Claude with templates to auto-create: for Email HQ, it scans 4 weeks of sent Gmail, extracts patterns (greetings like "Hey Name," signoffs, inbox zero workflow: 2-minute rule, labels, archive/snooze logic), and builds Email HQ/CLAUDE.md stacking on root voice rules. Result: Emails reference prior threads, follow conventions, sound like you.
For Personal Finances, upload 12 months of statements; Claude categorizes spending (e.g., Bumble Premium), builds Excel with tabs (Transactions, Yearly/Monthly Summary, Category Taxonomy), and remembers corrections (e.g., "Canva is subscriptions, not freelancers"). Project subfolders (e.g., mortgage refinance under Housing) inherit the same structure.
Build 2-3 workstations first; expand as needs arise. Use cases: Route screenshots to copywriting frameworks; post-meeting, auto-draft follow-ups pulling calendar/transcripts; create Notion projects (e.g., Boston trip July 17-24) filling properties/sections per your conventions.
Pro Tips: Session Audits and Token Optimization for Production Use
End sessions with "/session-audit" (custom skill from toolkit): scans conversation for unsaved principles/preferences, adds to memory.md. Keeps system evolving without manual updates.
Token savers: Reference external files instead of embedding; Sonnet for <3 interdependent steps; no repeated rules. After 30 workstations, author advises starting slow to master interactions. Free toolkit provides templates; paid Academy offers pre-built systems. Builds implied context (e.g., projects, style) for reliable outputs, per Google's AI Essentials learnings.