When the guy who built <a href="https://img.allwinchina.org/ai-tools/claude-code/" title="Claude Code review”>Claude Code shares how he actually uses it, people pay attention. Boris Cherny posted his terminal workflow on X last week, and the developer community has been chewing on it ever since. Jeff Tang, a well-known voice in the space, put it bluntly: if you’re not reading this, you’re falling behind.
What struck me most is how unglamorous it is. No fancy IDE plugins. No custom UI. Just a terminal, some tabs, and a lot of discipline. But the results sound almost absurd: one person operating with the output of a small engineering team. One X user said it “feels more like Starcraft” than coding. That’s not hyperbole when you see how it works.
Five agents, one commander
The headline grabber is that Cherny runs five Claude agents simultaneously in his terminal. He numbers his tabs 1 through 5 and relies on iTerm2’s system notifications to know when one needs attention. While one agent runs a test suite, another refactors legacy code, and a third writes documentation. He also keeps 5-10 Claude sessions open in his browser, using a “teleport” command to move sessions between local and web.
This isn’t just multitasking. It’s orchestration. And it fits neatly with what Anthropic President Daniela Amodei said earlier this week about “do more with less.” While OpenAI is building trillion-dollar data centers, Anthropic is proving that better coordination of existing models can give you exponential gains. I find that argument more compelling than the raw compute race.
Why he picks the slowest model
Here’s where Cherny goes against the grain. In an industry obsessed with latency, he uses Opus 4.5 with thinking enabled for everything. It’s the heaviest, slowest model Anthropic offers. His reasoning: “It’s the best coding model I’ve ever used, and even though it’s bigger & slower than Sonnet, 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.”
This is a critical point for anyone managing engineering teams. The real bottleneck isn’t token generation speed. It’s the human time spent correcting mistakes. Pay the “compute tax” upfront with a smarter model, and you avoid the “correction tax” later. I’ve seen teams burn weeks on prompt engineering with smaller models when they should have just used the expensive one from the start.
One file that makes AI learn from its mistakes
LLMs have amnesia. They don’t remember your coding style or architectural preferences between sessions. Cherny’s fix is elegantly simple: a single file called CLAUDE.md checked into the git repository. “Anytime we see Claude do something incorrectly we add it to the CLAUDE.md, so Claude knows not to do it next time,” he wrote.
When a human reviews a pull request and spots an error, they don’t just fix the code. They tag the AI to update its own instructions. Product leader Aakash Gupta called it “every mistake becomes a rule.” The longer the team works together, the smarter the agent gets. This is the kind of pragmatic solution that actually works in production, not some theoretical framework.
Slash commands for the boring stuff
The “vanilla” workflow everyone praises is powered by custom slash commands checked into the project. Cherny highlighted /commit-push-pr, which he uses dozens of times daily. Instead of typing git commands, writing commit messages, and opening pull requests manually, the agent handles it all.
This kind of automation is where the real productivity gains live. It’s not about AI writing complex algorithms. It’s about AI handling the tedious scaffolding that eats up 80% of a developer’s day. Cherny’s setup treats the AI like a senior engineer who never complains about grunt work.
What this means for the rest of us
I’ve been watching the AI coding space for years, and this workflow feels different. It’s not a demo. It’s not a polished product video. It’s one of the most capable engineers in the world showing exactly how he gets things done, warts and all. The fact that he uses terminal tabs and a markdown file instead of some bespoke tool is telling.
The takeaway for teams evaluating AI coding tools: don’t optimize for the wrong thing. Speed matters less than correctness. Orchestration matters more than raw power. And the best investment you can make is a shared file that captures your team’s collective knowledge.
Cherny’s thread is a masterclass in practical AI usage. It’s worth reading in full, but more importantly, it’s worth trying to replicate. Even running two agents in parallel with a CLAUDE.md file will change how you work. Start there.
Comments (0)
Login Log in to comment.
Be the first to comment!