Picture this: someone wakes up hung over, opens Claude, and casually asks it to review last month’s expenses, finalize a newsletter, and roast their Bumble Premium spending. Then goes back to bed. That was the opening scene of a video from Jeff Su, and my jaw was on the floor watching it.
The creator behind this walkthrough, Jeff Su, breaks down exactly how he built a Claude Cowork setup that runs his entire work life. The wild part? It’s all just plain text files. No code, no scary configs, no PhD required.
🧠 Why this matters
Most people treat Claude like a chatbot. You ask, it answers, you forget, repeat. The author’s approach flips that. He turns Claude into a persistent operating system that remembers your projects, writes in your voice, and routes tasks to the right context automatically.
The magic comes from two files: CLAUDE.md and memory.md. The first is a master rule book. The second is a notepad Claude updates between sessions. Together they give the AI something it usually lacks: continuity. And that continuity compounds, which is why Jeff says you need to start building this today, not next month.
🏗️ How the architecture works
The creator uses a three-level hierarchy that mirrors the US legal system. Stay with me, the analogy actually clicks.
- Level 0 (Root): Like the US Constitution. Rules apply across everything you do. Voice principles, memory rules, routing maps live here.
- Level 1 (Workstations): Like state laws. Each workstation handles one area of life (Email HQ, Personal Finances, Newsletter HQ) and stacks on top of root rules.
- Level 2 (Projects): Like city ordinances. Specific projects get their own folder with their own CLAUDE.md, memory.md, and resources.
When Claude writes an email, it reads the root voice file first, then layers email-specific rules on top. That stacking is what makes the output sound like the user instead of generic AI mush.
🛠️ The build steps Jeff walks through
- Create a parent folder called something like “Cowork OS” in Documents.
- Drop in three starter files: CLAUDE.md, memory.md, and a voice-principles.md inside a 00-resources subfolder.
- Open the folder in Claude and star it so every new session defaults there.
- Install Obsidian just to read the markdown files cleanly. The author admits he doesn’t even know how to use Obsidian beyond viewing files.
- Build the voice profile by pasting a prompt that tells Claude to scan your last 30 sent emails (or 5 writing samples) and extract patterns into voice-principles.md.
- Spin up workstations one at a time. Email HQ first, then maybe Personal Finances, then a Newsletter HQ. Don’t build all 30 at once.
The author shared a clever demo with personal finances. He dropped 12 months of credit card statements into Claude, and it built a master spending tracker spreadsheet with four tabs: transactions, yearly summary, monthly summary, and category taxonomy. When Claude miscategorized Canva as “freelancer,” he corrected it once. Next month it remembers.
💡 The use cases that sold me
The expert showed a few examples that hit different once you understand the system:
- Screenshot a copywriting framework, paste it in, say “figure out where this belongs.” Claude routes it to the right file.
- Say “draft a follow-up email from my last meeting.” Claude pulls the calendar event, finds the transcript, identifies the project, loads the right workstation, and drafts in the user’s tone.
- Say “I’m going to Boston July 17 to 24 for a wedding, set up a Notion project.” Claude fills in every property and section because it learned the user’s Notion conventions once.
🔋 Pro tips to keep token usage low
This is where the original poster saved viewers from rate-limit pain:
- Keep root CLAUDE.md under 300 lines. It loads on every session, so leanness matters. Use references to pull in extra files only when needed.
- Never repeat rules across files. If “no corporate jargon” lives in voice-principles, don’t restate it in Email HQ.
- Default to Sonnet, not Opus. Sonnet handles 80% of tasks at a fifth of the cost. Opus only kicks in when the task has three or more dependent steps.
- Run a session audit at the end. The author shares a starter skill that scans the conversation for unsaved preferences and writes them to memory before you close out.
My honest take after watching this twice: the system is genuinely simple, but the leverage compounds fast. Every correction you make once becomes a permanent rule. Every project you log becomes context for the next one. The person who starts this week is going to be miles ahead of someone who starts next month.
🚀 Worth watching in full
Jeff packs way more nuance into the video than I can fit here, including the routing map mechanics, the difference between universal and dedicated workstations, and the exact prompt templates he uses. If you’re serious about turning Claude into your second brain, the full walkthrough is the place to start. Check out the source link below for the complete breakdown.