When the creator of the world’s most superior coding agent speaks, Silicon Valley does not simply pay attention — it takes notes.
For the previous week, the engineering group has been dissecting a thread on X from Boris Cherny, the creator and head of Claude Code at Anthropic. What started as an off-the-cuff sharing of his private terminal setup has spiraled right into a viral manifesto on the way forward for software program improvement, with business insiders calling it a watershed second for the startup.
“If you happen to’re not studying the Claude Code greatest practices straight from its creator, you are behind as a programmer,” wrote Jeff Tang, a outstanding voice within the developer group. Kyle McNease, one other business observer, went additional, declaring that with Cherny’s “game-changing updates,” Anthropic is “on hearth,” probably dealing with “their ChatGPT second.”
The joy stems from a paradox: Cherny’s workflow is surprisingly easy, but it permits a single human to function with the output capability of a small engineering division. As one person famous on X after implementing Cherny’s setup, the expertise “feels more like Starcraft” than conventional coding — a shift from typing syntax to commanding autonomous items.
Right here is an evaluation of the workflow that’s reshaping how software program will get constructed, straight from the architect himself.
How operating 5 AI brokers directly turns coding right into a real-time technique recreation
Essentially the most placing revelation from Cherny’s disclosure is that he doesn’t code in a linear trend. Within the conventional “inner loop” of improvement, a programmer writes a operate, assessments it, and strikes to the following. Cherny, nonetheless, acts as a fleet commander.
“I run 5 Claudes in parallel in my terminal,” Cherny wrote. “I quantity my tabs 1-5, and use system notifications to know when a Claude wants enter.”
By using iTerm2 system notifications, Cherny successfully manages 5 simultaneous work streams. Whereas one agent runs a take a look at suite, one other refactors a legacy module, and a 3rd drafts documentation. He additionally runs “5-10 Claudes on claude.ai” in his browser, utilizing a “teleport” command handy off periods between the online and his native machine.
This validates the “do more with less” technique articulated by Anthropic President Daniela Amodei earlier this week. Whereas rivals like OpenAI pursue trillion-dollar infrastructure build-outs, Anthropic is proving that superior orchestration of current fashions can yield exponential productiveness features.
The counterintuitive case for selecting the slowest, smartest mannequin
In a stunning transfer for an business obsessive about latency, Cherny revealed that he completely makes use of Anthropic’s heaviest, slowest mannequin: Opus 4.5.
“I take advantage of Opus 4.5 with pondering for all the pieces,” Cherny explained. “It is the perfect coding mannequin I’ve ever used, and although it is greater & slower than Sonnet, since you must steer it much less and it is higher at instrument use, it’s nearly all the time sooner than utilizing a smaller mannequin ultimately.”
For enterprise expertise leaders, this can be a important perception. The bottleneck in fashionable AI improvement is not the technology velocity of the token; it’s the human time spent correcting the AI’s errors. Cherny’s workflow means that paying the “compute tax” for a better mannequin upfront eliminates the “correction tax” later.
One shared file turns each AI mistake right into a everlasting lesson
Cherny additionally detailed how his group solves the issue of AI amnesia. Commonplace giant language fashions don’t “bear in mind” an organization’s particular coding fashion or architectural choices from one session to the following.
To deal with this, Cherny’s group maintains a single file named CLAUDE.md of their git repository. “Anytime we see Claude do one thing incorrectly we add it to the CLAUDE.md, so Claude is aware of to not do it subsequent time,” he wrote.
This apply transforms the codebase right into a self-correcting organism. When a human developer opinions a pull request and spots an error, they do not simply repair the code; they tag the AI to replace its personal directions. “Every mistake becomes a rule,” famous Aakash Gupta, a product chief analyzing the thread. The longer the group works collectively, the smarter the agent turns into.
Slash instructions and subagents automate probably the most tedious elements of improvement
The “vanilla” workflow one observer praised is powered by rigorous automation of repetitive duties. Cherny makes use of slash instructions — customized shortcuts checked into the venture’s repository — to deal with complicated operations with a single keystroke.
He highlighted a command known as /commit-push-pr, which he invokes dozens of occasions every day. As an alternative of manually typing git instructions, writing a commit message, and opening a pull request, the agent handles the paperwork of model management autonomously.
Cherny additionally deploys subagents — specialised AI personas — to deal with particular phases of the event lifecycle. He makes use of a code-simplifier to scrub up structure after the principle work is completed and a verify-app agent to run end-to-end assessments earlier than something ships.
Why verification loops are the actual unlock for AI-generated code
If there’s a single motive Claude Code has reportedly hit $1 billion in annual recurring revenue so shortly, it’s possible the verification loop. The AI isn’t just a textual content generator; it’s a tester.
“Claude assessments each single change I land to claude.ai/code utilizing the Claude Chrome extension,” Cherny wrote. “It opens a browser, assessments the UI, and iterates till the code works and the UX feels good.”
He argues that giving the AI a option to confirm its personal work — whether or not by means of browser automation, operating bash instructions, or executing take a look at suites — improves the standard of the ultimate end result by “2-3x.” The agent does not simply write code; it proves the code works.
What Cherny’s workflow alerts about the way forward for software program engineering
The response to Cherny’s thread suggests a pivotal shift in how builders take into consideration their craft. For years, “AI coding” meant an autocomplete operate in a textual content editor — a sooner option to sort. Cherny has demonstrated that it may well now operate as an working system for labor itself.
“Learn this for those who’re already an engineer… and wish extra energy,” Jeff Tang summarized on X.
The instruments to multiply human output by an element of 5 are already right here. They require solely a willingness to cease pondering of AI as an assistant and begin treating it as a workforce. The programmers who make that psychological leap first will not simply be extra productive. They’re going to be taking part in a completely completely different recreation — and everybody else will nonetheless be typing.
