Managing a heavy B.Tech workload at NIT Allahabad while running a solo content channel felt like holding down two full-time jobs.
The bigger problem was the tools. Tasks lived in Notion. Scripts lived in Google Docs. Research happened in Perplexity. Drafting and planning happened in ChatGPT. Every time I opened a new AI chat, I was re-explaining my background from scratch, burning hours just to get back to where I’d left off.
I didn’t need another chatbot. I needed one connected system that could carry context across tools and pick up where I left off, every morning, without me rebuilding it by hand.
So I built my own “Cowork OS” in Claude. It now saves me 12 hours a week. Here’s exactly how I built it.
Resources
- Starter templates (copy-paste templates for CLAUDE.md, MEMORY.md, and Voice_Principles.md, coming soon)
- Obsidian (a free app to read and edit your Markdown files as a clean visual vault)
Why AI forgets you every morning
In my second year, I was using multiple tools:
- Notion for tasks, projects, and notes
- Google Docs for scripts and brain dumps
- Perplexity for deep research
- ChatGPT, Claude, and Gemini for brainstorming and editing
None of them talked to each other. So every new AI chat meant explaining who I am, who my audience is, how I write, and what I'm working on. Again.
I tried multiple methods to fix it. It broke fast, because every time a project changed, I had to edit dozens of prompts by hand. I didn't need better prompts. I needed an AI that remembers my context and works across my apps.
What Cowork OS actually is
Cowork OS is a folder on your computer full of Markdown (.md) text files. You point Claude Cowork at the folder. Claude reads the files at the start of every session, follows your instructions, and updates them as you work.
There are no databases and no code. If you can open a text document, you can build a Cowork OS.
The golden rule: rules vs. facts
The most important idea in the whole system is keeping rules and facts in separate files:
- CLAUDE.md holds rules, anything with "always", "never", or "format as".
- MEMORY.md holds facts that change, like project statuses, recent decisions, and contacts.
Rule of thumb: Ask, "Does this tell Claude how to behave, or does it record something that could change next week?" Behavior goes in CLAUDE.md. A changing status goes in MEMORY.md.
How the 3 levels stack
To keep Claude fast, accurate, and cheap to run, my Cowork OS has 3 levels of files.
1. The root
These files apply to every task:
- CLAUDE.md: my global tone, guardrails, memory rules, and a routing map.
- MEMORY.md: my bio, active projects, and core background facts.
- Resources/: heavy files like Voice_Principles.md that Claude loads only when needed.
2. 18 Areas
Each Area is one big part of my life, like Email, Personal Finances, Design Studio, Learning & Growth, or Travel. Every Area folder gets its own mini CLAUDE.md, MEMORY.md, and Resources/ folder.
3. Active projects
Short-term projects live inside an Area. My B.Tech major project sits inside Learning & Growth.
For example, when I ask Claude to draft an outreach email, it doesn't load rules for my whole life. It loads my root rules so it sounds like me, plus my Email HQ rules for formatting and sign-offs. The rules stack, and nothing gets written twice.
Pro tip: Give every task type exactly one home folder in your routing map. If two Areas overlap, Claude has to guess which rules to load, and the output gets worse.
Build your Cowork OS
To set this up:
- Create a folder called "Cowork OS" in your Documents.
- Inside it, create CLAUDE.md, MEMORY.md, and a folder called Resources.
- Open Claude Cowork, select the folder, click "Always Allow", and set it as your default.
1. Write your root CLAUDE.md
Keep it under 300 lines. Mine is 196. A shorter file saves tokens and stops Claude from lagging. Start with this template:
2. Fill in your MEMORY.md
Add 3 simple sections:
- About me: your role, college or job, and goals.
- Active projects: a bulleted list with statuses.
- Core memory: background facts that rarely change.
3. Teach it your writing voice
Paste 5 of your best writing samples into Claude with this prompt:
Save the output as Voice_Principles.md inside Resources/.
4. Add Areas and connect your apps
Create a subfolder for each main part of your life, like Email HQ or Design Studio. Each one gets its own CLAUDE.md, MEMORY.md, and Resources/ folder.
Then go to Claude Settings and connect Gmail, Google Calendar, Google Drive, Notion, Canva, and VidIQ. Any multi-step workflow you repeat can become a 1-click skill.
6 rules that keep it fast and accurate
- The rule vs. fact test: Rules go in CLAUDE.md. Changing facts go in MEMORY.md.
- The 300-line ceiling: Keep your root CLAUDE.md under 300 lines. Move specific instructions into Area files.
- Archive, don't delete: When MEMORY.md gets too long, ask Claude to compress old context into Archive.md.
- Pick the right model: Use Sonnet for 80% of daily tasks. Switch to Opus only for complex projects where the steps depend on each other.
- Confirm before writes: Make Claude ask before it writes to Notion or changes a local file.
- Run a daily session audit: End every session with /session-audit. Claude finds new rules and preferences in the chat and saves them to the right .md file.
Where the 12 hours come from
Here are 6 workflows running in my system right now:
- Morning planning: I type "Good morning, let's plan today." Claude reads my calendar, checks MEMORY.md for deadlines, and gives me a time-blocked plan. This used to take 20 minutes.
- YouTube scripts: I paste a rough voice brain-dump. Claude drafts the script in my voice, then creates a Notion page with Status, Platform, and Date already filled in.
- Email triage: Claude scans my 30+ daily emails, flags what matters, checks my calendar, and drafts replies in my voice for 1-click review.
- Research PDFs: I drop a 50-page engineering paper into Learning & Growth. Claude saves the core thesis, key formulas, and a 3-bullet summary as a Markdown file.
- Expenses: I drop my credit card statement into Personal Finances. Claude sorts every transaction, updates my spreadsheet, and warns me if a category goes over budget.
- Self-updating memory: After a long 30-minute session, /session-audit saves the preferences I worked out, so the next session starts with them.
You don't need all of this on day one. Build your root CLAUDE.md and MEMORY.md in 5 minutes, ask Claude a 1-line question to check it reads them, and add an Area only when the same need keeps coming back.
If you enjoyed this...
You might also like: What it's really like working in a Shark Tank India startup (coming soon)

