This is how Anthropic uses Claude Code internally. 10 tips from the creator, Boris Cherny: In January, his top tip was: "Most sessions start in Plan mode." I personally used it (almost) every day. In June he said new models need no planning step. So he runs auto mode for everything (a second model checks each action for him). Last week he shared how he prompts instead. You tell Claude 3 things: 1. What you want it to do 2. How much effort to spend 3. How to check it did the right thing "There's no secret to prompting," he says. You talk to Claude like a coworker. (It's really that simple now.) Here are all 10, and what I'd type: 1. A loop prompts Claude for him → /loop 2. The app makes his worktrees → /desktop 3. His routines babysit every PR → /schedule 4. He builds from his phone → /remote-control 5. He keeps prompts and tools minimal → /context 6. Mistakes become rules → CLAUDE.md or a skill 7. He skips the plan, runs auto mode → Shift + Tab 8. He prompts like a coworker → goal, effort, check 9. Claude checks itself → "Check it works in Chrome" 10. He runs every agent from one tab → claude agents But his biggest idea didn't change. He puts every mistake into CLAUDE.md or a skill. Then, in his words, Claude "can just run forever". Which change are you making first?