So over the weekend, Boris Cherny, the creator of Claude Code, tweeted out how he's using Claude Code in his vanilla setup.

So over the weekend, Boris Cherny, the creator of Claude Code, tweeted out how he's using Claude Code in his vanilla setup. It's anything but vanilla, and it blew up on Twitter. So let's run through how he uses Claude Code so we can all learn from it together. Now when I read this list, I was actually surprised I'm doing like one or two of the things he's doing. So first thing, he's using agents for everything. He is always deploying agents for all of his tasks. Number two, he's able to deploy those agents successfully and with the right context and guidance because he's using plan mode for all of his prompts. So he's going into plan mode, creating a structured plan with Claude, and then spin it off into multiple agents. Three, he's using Claude web and the Claude app to spin off new web agents way more than I thought anybody was. And a lot of times what he's doing is he's planning out his work, spawning agents, and then shipping off more work to those web agents with the ampersand, with the and sign, and just activating them there. Number four, we're not using enough terminal tabs. He has got five to 10 Claude tabs in his terminal window at any given time, all running in parallel, all in the same repo. So he's planning out in five different tabs, activating agents within those five tabs, and spinning off additional agents in the web version with the ampersand sign. Five, him and his team at Claude are using the claude.md file as their source of truth, as their holy bible, so to speak. So their whole team is contributing to the claude.md file. They are contributing to it during PRs, they're having Claude write back to it. Everything they learn, every advancement they make and how they work together is going in that MD file and to the rest of them so it standardizes the work they're doing. Six, and this is a new one for me, slash commands. He's running a bunch of slash commands. He's making them with Claude and then running them almost like mini skills where you just hit a slash command, commit and PR, and it knows exactly how to commit it and PR it. I would have thought that was a skill. Seven, instead of having it run dangerously, where a command you can use to run it at the start and Claude will run without asking any permissions, he is going into the permissions and essentially turn on all the permissions and the requests that Claude will make, knowing what he knows about the repo and their setup, that he trusts it to run those commands, and he's leaving the other ones unchecked. And number eight, he's giving Claude this really foundational and important feedback loop so it can test its own work, validate what it's doing, and come back knowing that it's done it right. So using an MCP, using Claude Chrome to test in the web browser, giving Claude that feedback loop to know when it's done and know that it's been tested and validated. I see an opportunity to load all of these into my workflow. Which ones are you gonna focus on first?