AI CASE
An AI OS for my own work
An AI model remembers nothing between two conversations, and I work across domains that must not mix. I built a workplace in which Claude knows those domains, picks up work where the previous session stopped, and follows rules I do not have to explain again every time.
Published on
- PERIOD
- 2026 - heden
- USED
- ClaudeObsidianGitHubn8n
The problem
A model remembers nothing between two sessions. Because next to my job I run a business of my own, coach at a gym and work for a client, every conversation started with explaining again who I am, what I am working on and which information must not come together.
The approach
I started by writing things down. My work environment is written up so that a model can work in it on its own, with its own context per domain and a wall between work and private life that is guarded automatically. Work that is not finished, I hand over to the next session, and what I know sits in a second brain Claude can search. Rules that must always hold are automatic checks. On top of that run agents that attack a plan before it is carried out and check work before it is called done.
How it went
Step 01 ·
One workplace
Two separate AI assistant setups merged into one workplace, with one instruction file that says who I am and how I work.
Step 02 ·
A profile per domain
Every domain got its own context, and at the start of a session the question decides which domain leads. Later a wall between work and private life was added, guarded automatically.
Step 03 ·
Handing work to the next session
Work that is not finished, I write down so a new session can pick it up cold. A handover only closes once every open item is proven done.
Step 04 ·
A second brain
My knowledge got a layer of its own that Claude can search, first for one domain and two days later for all my private domains.
Step 05 ·
A rule becomes a check
A writing rule moved from the instructions to an automatic check that stops the mistake before it is saved. Since then that is the form for every rule that must always hold.
Step 06 ·
From assistant to agents
First an agent that attacks a plan before it is carried out, then one that checks whether work is really done, and finally a layer that goes through the open work per domain at night. In the morning I decide what happens to it.
Step 07 ·
Autonomy measured
I started measuring what share of the work lands without me, with a fixed definition of what counts.
The result
A new session picks up work where the previous one stopped, without me explaining who I am or which domain we are in. A mistake against a fixed rule is stopped before it is saved, and at night agents go through the open work per domain. Most of the work still runs through me, and I now measure that.
What went wrong
A rule written in the instructions is not followed by itself. A writing rule was in there from the start and was still broken, and only after the second breach did I turn it into a check that stops the mistake. Now I ask of every new rule first whether it must always hold, and if it must, it becomes a check.