Claude Code’s Origin Story: Why the Real Edge Isn’t the Model

Anthropic quietly launched a page called “The Making of Claude Code” that even lets you Read in Terminal . The author pieced together a year‑long narrative of Claude Code’s development, drawing from recent interviews with its creator Boris Cherny.

A two‑like "toy"

In late 2024 Boris joined Anthropic Labs, a tiny internal incubator, motivated by what he called a "product overhang" – models were far ahead of any product that could fully exploit them. At that time the most advanced coding assistance was simple tab‑completion (Sonnet 3.5). Believing models would soon generate whole code blocks, he built Claude CLI, announced it internally, and received only two likes, with many doubting a terminal‑only tool.

The moment of conviction

Boris asked the model a trivial question – "What music am I listening to?" – and the model invoked bash to inspect system processes, retrieve the player status, and answer correctly. This emergent tool‑use capability convinced him that the model could think beyond scripted prompts.

Why the terminal?

Initially the terminal was chosen out of necessity – Boris was the sole developer and it was the quickest way to prototype. Later the team deliberately kept the minimal shell because a lightweight interface could keep pace with the rapid iteration of models, allowing the product to capture new model capabilities immediately.

Six months of silence and the May breakthrough

Claude Code was publicly released in February 2025 but was "almost unusable" for the first six months; even Boris used it for only about 10 % of his code. The turning point came in May 2025 with the launch of Opus 4, after which each subsequent model upgrade (Opus 4.5, 4.6, 4.7) lifted the product’s growth curve. Boris quoted the team’s philosophy: they built for the "next‑generation" model even though there was no immediate product‑market fit.

Leaked code: TypeScript + React

The leaked repository showed a straightforward stack – TypeScript and React – chosen because those languages dominate the model’s training data, making generated code easier for the early‑stage model.

Productivity explosion

From October/November 2025 Boris stopped hand‑writing code entirely, delegating to Claude Code. He began submitting dozens of PRs daily, peaking at 150 in a single day, using the Claude app on his phone with multiple sessions and agents. He described a "loop" workflow where Claude repeatedly runs tasks such as auto‑fixing CI, maintaining CI health, and scraping Twitter feedback every 30 minutes.

The memory‑leak story

A junior engineer handed a memory‑leak debugging task to Claude Code. Claude captured a heap snapshot, wrote an analysis tool, pinpointed the issue, and submitted a PR faster than Boris could with traditional debugging tools.

A brief detour to Cursor

Boris briefly left Anthropic for Cursor, attracted by product and team, but returned after two weeks because he missed Anthropic’s "safety‑first" mission.

Organizational advantage

According to Boris, the real edge isn’t the model – Anthropic uses the same model as the outside world – but the organization’s structure and processes. The company dogfoods Claude Code, writes no hand‑coded SQL, and even lets AI agents coordinate via Slack, creating a fully AI‑driven development pipeline.

Printing press metaphor

Boris likens today’s AI wave to the 15th‑century printing press: just as literacy exploded after cheap books, software development will become ubiquitous, with domain expertise becoming the scarce skill. He predicts the "software engineer" title may fade, replaced by interdisciplinary builders.

Key takeaways

Don’t hard‑wire models : give them tools and goals; scaffolding only adds 10‑20 % benefit, while the next model erases that gap.

Build for models six months ahead : immediate PMF is less important than being ready when a stronger model arrives.

Move fast and watch misuse : today’s side projects (e.g., restoring wedding photos, analyzing MRI) often point to tomorrow’s core products.

In the end, the real problem solved is the repetitive act of typing, not the conceptual work of deciding what to build.