Claude Code: Open Permissions, Hard Limits
I recently started using Claude Code but I found the permission prompts both hard work and unreliable.
Trying to work out what I was agreeing to was hard work - and all too often the agent managed to do things I hadn’t wanted.
So I turned the inbuilt soft limits off (--dangerously-skip-permissions) and put Claude in a container I call a devbox. Inside the devbox Claude can do whatever it likes. The container decides what can actually be written, where it can connect to, and which credentials it holds, and Claude can’t change any of that from inside.