For months, I operated under a simple belief: if I write good enough instructions, the AI agent won't break anything.
I was wrong.
The model will occasionally:
- Misinterpret what "don't touch" means
- Lose track of which files are safe to edit
- Decide to "clean up" something you didn't ask it to
- Run a refactor script that goes one folder too deep
Every time, the prompt looked reasonable. Every time, something still went wrong.
The realization
Prompts are requests. Operating system boundaries are enforcements.
Requests can be ignored. Enforcements cannot.
So I stopped trying to write the perfect prompt and started giving each AI agent its own isolated drive letter on Windows.
How it works
Create a folder for the task, then map it to a drive:
subst Z: C:i-workspaces\project-alpha
Point your AI tool at Z:\. The agent can read, write, delete, and do whatever it wants inside that drive — but it literally cannot see anything outside of it.
This isn't a suggestion. It's not a permission setting in the AI tool. It's the Windows filesystem saying "this drive only exists here."
The tools I use
I use ScsSubst (https://subst.scsoi.com/) for this. It's a GUI wrapper around the Windows subst command that lets you create persistent drive mappings that survive reboots. $9.99 lifetime, one-time purchase.
For simple cases, the built-in subst command works fine. But ScsSubst adds:
- Persistent mappings across reboots
- GUI to manage all your mapped drives
- Shows real disk usage
- One-click cleanup when you're done
The mental shift
I used to think of AI safety as a prompt engineering problem. Now I think of it as an operating system problem.
You wouldn't run an untrusted program as admin and hope it behaves. Why are you running an AI agent on your entire home directory and hoping it behaves?
Give it its own drive. It's the cheapest, most reliable sandbox you can build for $10.
Top comments (0)