The tool that kickstarted the vibe coding revolution was, itself, vibe coded. Around 90% of Claude Code's codebase was written by Claude Code. The head of the project says 100% of his recent contributions are written through it. If that doesn't make you pause and reconsider what "software engineering" means in 2025 and beyond, nothing will.

It started with a music player

Claude Code didn't begin as a grand product vision. In late 2024, Anthropic engineer Boris Cherny was tinkering with Claude in the terminal, building a small prototype that could tell him what music he was listening to at work via AppleScript. It was a cool hack, but not much more than that. Then Cherny gave the prototype access to the filesystem and bash commands. Suddenly, Claude could read files, follow imports, and explore an entire codebase on its own. Cherny described this as discovering "product overhang," the model already had the capability, but no product was built around it. Within two months, they had a dogfooding-ready version. On day one of internal release at Anthropic, 20% of the engineering team started using it. By day five, that number hit 50%.

Designed so Claude could build itself

Here's where the recursion gets interesting. The Claude Code team deliberately chose a tech stack that Claude was already good at: TypeScript, React (with the Ink framework for terminal UIs), and Bun for packaging. In AI terminology, they picked technologies that were "on distribution" for the model. Boris Cherny explained the reasoning: "We wanted a tech stack which we didn't need to teach, one where Claude Code could build itself. And it's working great. Around 90% of Claude Code is written with Claude Code." This wasn't an accident or a marketing stunt. It was a deliberate architectural decision. The team tries to write as little business logic as possible, keeping Claude Code as a lightweight shell on top of the model. Every time a new model version ships, they delete code rather than add it. With the 4.0 models, they removed around half the system prompt because the model no longer needed it.

The numbers back it up

By January 2026, Cherny told WIRED that 100% of his personal code contributions over the previous two months had been written through Claude Code. Not assisted by it, but written through it. For the broader Claude Code team, the figure sits around 95%. The product hit $500 million in annualized recurring revenue by mid-2025. By November 2025, that number crossed $1 billion. At any given moment, Cherny says he has five to ten Claude Code agents running in parallel across his terminal, phone, and browser. And when Anthropic doubled its engineering headcount, pull request throughput per engineer didn't dip (as typically happens during rapid hiring). It went up 67%, credited largely to Claude Code adoption.

What "vibe coded" actually means here

The term "vibe coding" was coined by Andrej Karpathy in early 2025 to describe writing software by describing what you want in natural language and letting AI handle the implementation. You focus on the vibe of what you're building, not the syntax. Claude Code is one of the most popular tools for this workflow. It runs in your terminal, reads your codebase, edits files, and executes commands. Developers describe a feature or bug fix in plain English, and Claude handles the implementation, testing, and iteration. The meta twist is that this exact workflow is how Claude Code itself gets built. The team prototypes aggressively, sometimes building 20 variations of a feature in two days, each one prompted in natural language and generated by Claude. When Boris Cherny built the todo list feature, he went through roughly 20 prototypes by writing short, conversational prompts and letting Claude generate each version.

The simplicity principle

One of the most counterintuitive aspects of Claude Code's architecture is how simple it is. There's no virtualization, no Docker containers, no cloud sandboxing. It runs locally, reads your filesystem directly, and executes bash commands on your machine. The team's guiding principle is to always pick the simplest possible option. The most complex piece is the permissions system, which asks before executing potentially dangerous actions and supports per-project, per-user, and per-company settings. Beyond that, the product philosophy is to get out of the model's way. Cherny put it bluntly: "When you look at a lot of coding products, they get in the way of the model. They add scaffolding by adding UI elements and other parts that clutter things, so that the model running in those tools feels like it's hobbling on one foot."

The recursive future

There's something philosophically striking about a tool that builds itself. Claude Code is not just a product that uses AI, it's a product that is AI in a very literal sense. The model writes the code. The model debugs the code. The model prototypes new features. The humans guide, review, and decide what ships. This pattern is likely to accelerate. As models get more capable, the percentage of AI-written code in AI tools will only grow. The tools we use to build software are increasingly built the same way we use them: by describing what we want and letting the machine figure out how. The vibe coding tools were vibe coded. And if the trend continues, the next generation of these tools will be even more so.

References