Boris Cherny builds Claude Code at Anthropic. Last week he posted his personal workflow on X, and the engineering community basically stopped what it was doing to study every detail. What sounds like a casual "here's my setup" thread turned into one of the most discussed posts in the dev world this year.
The big reveal? Cherny doesn't code the way most people think. He runs five instances of Claude simultaneously in his terminal, numbering his tabs 1 through 5 and using system notifications to know when any agent needs his attention. While one agent runs tests, another refactors code, and a third writes documentation. He also keeps 5 to 10 more Claude sessions running in his browser, with a "teleport" command that lets him move work between the web and his local machine. As one user put it, it "feels more like Starcraft" than traditional programming.
Here's the part that surprised a lot of people. In an industry obsessed with speed, Cherny exclusively uses Opus 4.5, Anthropic's heaviest and slowest model. His reasoning is practical. Even though it takes longer per response, you spend far less time correcting mistakes and steering the model back on track. In his words, "since you have to steer it less and it's better at tool use, it is almost always faster than using a smaller model in the end." The real bottleneck was never token generation. It was always the human time spent fixing AI errors.
One of the smartest details in the thread is how his team handles AI memory, or the lack of it. They maintain a single file called CLAUDE.md in their git repository. Every time Claude does something wrong, they add a note to that file so the model knows not to repeat the mistake. As product leader Aakash Gupta observed, "every mistake becomes a rule." The longer a team works this way, the sharper the agent gets at following their specific standards.
Cherny also leans hard on automation for the boring stuff. He uses slash commands, which are custom shortcuts stored in the project repo, to handle repetitive operations. One command called /commit-push-pr handles git commits, commit messages, and pull request creation in a single keystroke. He runs it dozens of times a day. On top of that, he deploys specialized subagents for tasks like simplifying architecture after the main work is done and running end-to-end verification before anything ships.
The verification loop might be the most important piece of the whole setup. Cherny says Claude tests every change he lands using a Chrome extension that opens a browser, tests the UI, and iterates until things actually work and feel right. He estimates that giving the AI a way to verify its own output improves final quality by 2 to 3x. The agent doesn't just generate code. It proves the code works.
Why this matters for anyone building with AI tools right now: the gap between "using AI to autocomplete code" and "orchestrating AI agents like a fleet" is widening fast. Cherny's workflow shows that the multiplier isn't coming from better models alone. It comes from better systems around the models. A single file for memory, parallel agents, verification loops, and automation of version control grunt work. None of it is exotic technology. It's just disciplined orchestration.
The developers paying attention are already adopting pieces of this. The ones who aren't may find themselves wondering why a solo builder with five terminal tabs is shipping faster than their entire team.
