When the creator of the world's most advanced coding agent speaks, Silicon Valley doesn't just listen — it starts taking notes.
Over the past week, engineers have been poring over a thread on X from Boris Cherny, the creator and lead of Claude Code at Anthropic. What started as a casual look at his personal terminal setup has quickly evolved into a viral blueprint for the future of software development, with insiders calling it a defining moment for the startup.
"If you're not reading the Claude Code best practices straight from its creator, you're behind as a programmer," wrote Jeff Tang, a well-known voice in the dev world. Kyle McNease, another close observer, went even further, arguing that with Cherny's "game-changing updates," Anthropic is "on fire" and may be approaching "their ChatGPT moment."
The buzz comes from a paradox: Cherny's setup looks almost minimalistic, yet it lets one engineer operate with the throughput of an entire team. As one X user remarked after adopting his approach, the workflow "feels more like Starcraft" than conventional programming — shifting from typing code to directing autonomous units.
Below is a breakdown of the workflow that is quietly redefining how software is built, as described by its architect.
How five parallel AI agents turn coding into a real-time strategy match
The most eye-opening detail from Cherny's post is that he doesn't work through tasks sequentially. In the classic development "inner loop," a developer writes a function, tests it, then moves on. Cherny instead behaves like a commander overseeing multiple fronts.
"I run 5 Claudes in parallel in my terminal," Cherny wrote. "I number my tabs 1-5, and use system notifications to know when a Claude needs input."
Using iTerm2 system notifications, he orchestrates five concurrent workstreams. One agent might be running tests, another refactoring a legacy component, a third drafting documentation. On top of that, he keeps "5-10 Claudes on claude.ai" open in his browser, using a "teleport" command to seamlessly transfer sessions between the web and his local environment.
This approach reinforces the "do more with less" philosophy outlined by Anthropic President Daniela Amodei earlier in the week. While rivals like OpenAI chase massive infrastructure build-outs, Anthropic is showing that smarter coordination of existing models can unlock outsized productivity gains.
Why picking the slowest, smartest model can actually be faster
In a move that runs counter to an industry obsessed with speed, Cherny shared that he relies exclusively on Anthropic's largest and slowest model: Opus 4.5.
"I use Opus 4.5 with thinking for everything," Cherny said. "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."
For CTOs and engineering leaders, this is a key takeaway. The real bottleneck in AI-assisted development isn't token generation speed; it's the human time spent debugging and correcting. Cherny's method implies that paying a higher "compute tax" upfront for a more capable model reduces the "correction tax" later on.
How a single shared file turns every AI error into institutional memory
Cherny also explained how his team tackles the issue of AI forgetfulness. Out of the box, large language models don't retain a company's conventions or architectural choices across sessions.
To fix this, his team keeps a single file called CLAUDE.md in their git repo. "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.
This turns the codebase into a continuously learning system. When a developer reviews a pull request and finds an issue, they don't just patch the code; they also update the AI's playbook. "Every mistake becomes a rule," observed Aakash Gupta, a product leader who broke down the thread. Over time, the agent effectively levels up alongside the team.
Slash commands and subagents strip away the drudgery of development
The "vanilla" workflow that one commenter praised is actually powered by aggressive automation of routine work. Cherny relies on slash commands — custom shortcuts stored in the repo — to compress multi-step operations into a single action.
He calls out a command named /commit-push-pr, which he triggers dozens of times a day. Instead of manually running git commands, crafting a commit message, and opening a pull request, the agent handles the entire version-control pipeline automatically.
He also uses subagents — focused AI personas — for distinct stages of the lifecycle. A code-simplifier agent tidies up architecture once the main implementation is done, and a verify-app agent runs end-to-end checks before changes go live.
Why verification loops are the real breakthrough for AI-written code
If there's one core reason Claude Code has reportedly reached $1 billion in annual recurring revenue so rapidly, it's likely the verification loop. The AI isn't just generating text; it's actively testing its own output.
"Claude tests every single change I land to claude.ai/code using the Claude Chrome extension," Cherny wrote. "It opens a browser, tests the UI, and iterates until the code works and the UX feels good."
He contends that giving the AI tools to validate its work — via browser automation, bash commands, or test suites — boosts final quality by "2-3x." The agent doesn't merely write code; it demonstrates that the code behaves correctly.
What Cherny's setup reveals about where software engineering is headed
The response to Cherny's thread points to a deeper change in how developers view their role. For a long time, "AI coding" meant smarter autocomplete in an IDE — essentially a faster keyboard. Cherny has shown it can now operate as an orchestration layer for human labor.
"Read this if you're already an engineer… and want more power," Jeff Tang wrote on X.
The capability to multiply an individual's output by five is already available. The missing piece is the mindset shift: stop treating AI as a sidekick and start managing it as a distributed workforce. The developers who adopt that frame first won't just ship more. They'll be operating in a different paradigm — while everyone else is still just typing.