Okay y’all, this one is wild.
Boris Cherny, the guy who built Claude Code, shared his full workflow in a long post. And let's just say, the developer community is currently having a collective existential crisis. The good kind.
But first let’s talk numbers, because they’re absolutely mind blowing:
259 pull requests in 30 days
497 commits
40,000 lines added
38,000 lines removed
Every. Single. Line. All written by Claude Code with Opus 4.5. Apparently, Opus 4.5 needs less hand-holding, is better at tool use, and finishes faster overall.
And here’s the kicker: his setup is surprisingly vanilla.
No exotic hacks. No galaxy-brain tricks. Just a brutally disciplined workflow that turns one human into an entire engineering department.
Here’s the setup (Yes, It’s Insane)
Cherny runs:
Five Claudes in parallel in his terminal
Another 5 to 10 Claudes on claude.ai in his browser
Use system notifications to know when a Claude needs input
Uses a custom “teleport” command to move sessions between web and local environments.
And yes, he kicks off sessions from his iPhone throughout the day
He starts in Plan Mode (Shift + Tab twice), iterates until the plan is airtight, then switches to auto-accept edits.
From there, almost everything is automated.
Slash commands handle the busywork. His /commit-push-pr command, which he uses dozens of times a day, stages changes, commits code, pushes updates, and opens pull requests automatically.
Then come the specialized AI personas, also called subagents.
Each one handles a specific task:
A code-simplifier cleans up after the main work
A verify-app agent runs end-to-end tests before anything ships
It’s basically AI managing AI.
Think about that for a second.
While one agent runs tests, another refactors code, a third writes documentation, and a fourth debugs. At this point, he’s not coding anymore. He’s orchestrating.
One developer put it perfectly: “It feels more like StarCraft than traditional coding.”
But what really has people talking is this: Cherny’s team maintains a single shared file in their repo called CLAUDE.md.
Every time Claude makes a mistake, they add it to the file. During code reviews, they literally tag @claude and ask it to update its own instructions.
The result? Every mistake becomes a rule, the system gets smarter over time and the codebase learns from itself.
Seriously, it’s a self-correcting workflow where the longer you work with it, the better Claude performs.
Now here’s Cherny’s most important insight: Give Claude a way to verify its own work, and quality jumps 2 to 3×.
Claude doesn’t just ship code and hope for the best. For every change to claude.ai/code, it opens Chrome, tests the UI, and iterates until everything actually works.
In other words, it doesn’t just write code. It proves the code works. And if there’s one reason Claude Code hit $1B in revenue this fast, it’s probably this..
Why This Matters
This isn’t about working harder. It’s about working differently.
The future of software isn’t one developer writing more code. It’s one developer commanding systems that plan, build, test, verify, and improve themselves.
So the question is no longer, “Will AI change how we work?” It already has. The real question is: Are you using AI like a tool… or like a team?
Take a look at the detailed workflow here.
