Case study | AI workflows · Knowledge systems · Process design
WorkOS, my personal operating system
WorkOS is the personal operating system I built to work with AI: Markdown files, Claude skills and Notion setup that give it context and memory, so every session starts informed instead of from zero.
- Tools
- Markdown files, Claude, Notion, Obsidian
- My role
- Designer, builder and user
- Worked with
- Claude, as a build partner
The problem
AI assistants forget stuff. Every new session meant explaining my background, projects and preferences again, and the long prompts that carried all that context used a lot of tokens.
A related problem was visibility. I manage multiple projects and opportunities at once, and there was no single place that showed the status of each one.
What I found
Memory has to live outside the chat
If context is saved in plain files that I own and can edit in Obsidian, any session can pick it up. Nothing depends on what the AI happens to remember.
Load only what the task needs
Giving the AI everything every time is slow and costly. A map that points to the right file keeps sessions light and focused.
Rules, facts and status are different things
When they are mixed in one place, instructions pile up, repeat and start to conflict. Each kind needs its own home.
How it works
The setup is adapted from AI and organizational best practices and shaped around how I work. A root rulebook holds my standing instructions and a routing map that tells the AI which workstation folder to open for each kind of request, such as career, email or travel.
A live memory file acts as an index: active projects, current decisions and open loops, each pointing to the one file where the detail lives. Every fact or rule has a single home, and a simple test decides where new information goes: behavior goes in rules, lasting facts go in reference files, and status goes in memory.
Projects sit inside their workstation and inherit its rules. This website is one of them, with its own notes, and it is built and updated with AI through a review step before anything goes live.
What I built
- Context re-explained at the start of every session
- Long prompts and heavy token usage
- Instructions that repeated or conflicted
- No single view of the opportunities I was managing
- Every session starts with full context
- A routing map that loads only the relevant files
- One home for each rule, fact and status
- A session audit before wrapping up, so corrections and decisions are saved
- A periodic review of memories and rules to cut redundancy and conflicts
- A Notion CRM that tracks multiple opportunities in one view
The Notion CRM came out of this setup. I built it myself, with AI assistance, and it gives me control and visibility over all the opportunities I'm managing at any given moment.
Results
Memory loss and heavy token usage stopped being problems. Every session starts with full context, and the session audit makes sure everything is recorded before I close it.
It saves me a lot of time, and it is an ever-evolving process: I keep refining it as my work changes.