The way claude code and codex implemented admin permissions is fascinating. It clearly shows the differences in how these companies operate.

Claude Code asks you to explicitly invoke --dangerously-skip-permissions when starting claude. I regularly make mistakes when typing such a long string. This makes me conscious for an extra few seconds about what I am doing with skipping permissions.

Codex on the other hand has 2 commands that can do this, and the 2nd one is more interesting. There’s --dangerously-bypass-approvals-and-sandbox which reflects the side of openai which named their flagship product chatGPT.

The 2nd command however is --yolo which some growth guy clearly came up with (kudos, btw). This reflects the side which came up with ghiblify to make chatGPT images go viral.