The rules I set myself
Claude is fast and capable, so the guard rails matter more, not less: when it may act on its own, when it must ask, and what never goes through the chat.
Why rules
A fast assistant makes fast mistakes. Instead of watching every keystroke, I set a few rules once, write them down in memory so they survive between sessions, and let the work run inside them.
The rules
- Commit and push only when I ask. Claude prepares the change and shows it; publishing is my call, every time.
- Changes to my network and servers are shown first. Before touching a router, a firewall or a shared server, Claude lists the exact commands, the risk and the way back, and waits for a yes. When several changes are proposed together I often approve only the one that is needed.
- Secrets never go through the chat. Passwords and tokens are pasted by me directly where they live, or copied from machine to machine without being printed. To check one, Claude looks at its length or format, not at its value.
- Check before and after. On the shared server, a script verifies the network and the containers before a change and again after it. If the second run complains, Claude stops and tells me instead of trying to fix it.
- Say what was verified. Every answer ends with what was actually tested and what was not (“built and loaded in the browser, not tried on a real phone”), and with the decisions taken on my behalf.
- Downloads and installs need a yes, with name, source and size.
- Public pages stay generic. Nothing about my accounts, family or home infrastructure goes on this site; the details live in private memory.
What I would tell anyone starting
Write the rules down the first time you hit the problem, with the reason. A rule with a reason is applied correctly in situations you did not foresee.