According to VentureBeat, Boris Cherny, the creator and head of Claude Code at Anthropic, recently shared his personal AI coding workflow on X, sparking a viral reaction in the developer community. The thread revealed he runs five Claude instances in parallel in his terminal, exclusively uses the slowest but smartest Opus 4.5 model, and maintains a shared CLAUDE.md file to teach the AI from past mistakes. Industry observers like Jeff Tang and Kyle McNease declared this a potential “ChatGPT moment” for Anthropic, with Claude Code reportedly hitting $1 billion in annual recurring revenue. The workflow transforms coding from a linear task into a command-and-control operation, with Cherny comparing it to managing a fleet or playing a real-time strategy game like Starcraft.
The Commander, Not The Coder
Here’s the thing that’s really blowing minds. It’s not about the AI writing a line of code for you. It’s about redefining the entire job. Cherny isn’t sitting there typing. He’s orchestrating. Running five agents at once, using system notifications to know when one is stuck? That’s a completely different mental model. You’re no longer in the weeds of syntax. You’re managing a portfolio of work streams. One agent‘s testing, another’s refactoring, a third is writing docs. Your job is high-level direction and quality control. It sounds powerful, but man, does it also sound chaotic. How do you maintain context across all those threads? What’s the cognitive load like? I think the real insight is that we’ve been using AI to amplify our hands, but Cherny is using it to amplify his *attention*.
The Slow Model Paradox
Cherny’s insistence on using the biggest, slowest model, Opus 4.5, is a brilliant counter-narrative. In a tech world obsessed with shaving milliseconds off latency, he’s arguing for a “compute tax” upfront to avoid a massive “correction tax” later. And you know what? It makes total sense. A smarter model gets it right the first time more often, requires less hand-holding, and uses tools more effectively. You’re trading raw token generation speed for a higher-quality, more autonomous output. For businesses, this is huge. The bottleneck isn’t how fast the text appears on screen; it’s the human hours spent debugging and re-prompting. This is a lesson that probably applies far beyond just coding assistants.
The Hidden Risks of Scale
But let’s pump the brakes for a second. This sounds amazing, but what are we glossing over? First, the CLAUDE.md file. Turning every mistake into a rule is clever, but that file is going to become a monstrous, contradictory bible. Who curates it? What happens when rules conflict? It’s technical debt of a whole new flavor—prompt debt. Second, running a swarm of agents is a recipe for chaos if you’re not a supremely organized thinker like Cherny seems to be. For every developer who turns into a fleet admiral, how many will just create a tangled mess of half-finished, AI-generated code? And third, verification is key, but having the AI test its own UI work feels a bit like the fox guarding the henhouse. You still need that human in the loop, if only to ask, “Does this actually make sense for a user?”
A Different Game Entirely
So what does this mean? The hype is real, but it’s pointing at a real shift. The Claude Code platform isn’t just a fancy IDE plugin. It’s an attempt to build an operating system for software labor. The slash commands, the subagents, the verification loops—they’re creating a framework where the human provides strategy and the AI handles tactics and execution. The programmers who master this won’t just be 5x faster. They’ll be operating in a different dimension. Everyone else will still be typing. But look, this is still early days. It requires a massive shift in mindset and a tolerance for a new kind of complexity. The tool is here. The question is, are most developers ready to stop being coders and start being commanders?
