Been running coding agents a lot lately for a side project and had a realization I can't shake. The question I kept asking was "is this thing good at writing code." But that's not the one that protects me. The one that matters is "what will it refuse to do without asking." Because I'm not reading every change it makes, the stuff it won't do on its own is basically my safety net. Claude Code just added a refusal list for destructive commands, which made me realize I'd neve...