How Was Claude Code Built? Inside Anthropic's 'The Making of Claude Code' Oral History

Anthropic published “The Making of Claude Code” on July 6, 2026 - a feature story telling the inside story of how its coding agent came to be, through the words of the people who built it¹². Structured as a seven-part oral history, it collects testimony from co-founder Ben Mann, Claude Code head Boris Cherny, the development team, and early users, covering everything from the origins to the road ahead¹. The interviews were recorded between February and May 2026¹.

Claude Code is now a major player in the AI coding tools market, but the piece reveals that its starting point was a demo almost nobody at the company noticed. This article walks through the story that emerges from the testimony, and what it suggests for readers building or adopting AI products.

It Started With a VS Code Extension and an Internal Tool Called “clide”

According to Mann, the first product Anthropic ever built was, in fact, a coding assistant. It was a VS Code extension - a chat interface that returned four different suggestions for each prompt - and by spring 2022 it had around 100 external users¹³. In parallel, the research side had been thinking about autonomous software engineering since early 2022, building the original reinforcement learning codebase for training agents¹. On why coding was the focus, head of reinforcement learning Shauna Kravec explains: “we thought that the path to transformative AI would route through the ability to automate large chunks of software engineering work”¹.

What came next was an internal CLI tool called “clide” - a way to chat with Claude in the terminal for code editing and dev tasks¹³. Testimony describes it as wonky but “very, very ahead of its time”¹ - the elements of Claude Code were already visible.

From “Two or Three Likes” on Slack to a Real Product

The turning point came in 2024. Mann started the Labs team in January of that year, saying he “saw a hole in the market for agentic coding”¹. When Cherny joined Labs that September, he initially wanted to build something small - a linter - but Mann pushed back: “No, no, no, no, you have to do the big thing”¹.

Cherny recalls watching clide write an entire pull request: “I had never seen anything like that. It was just shocking. It felt like the future”¹. In about two days he built a demo he called “Claude CLI” and posted it on Slack. The reaction? “I think I got two or three likes”¹. But the very next day a colleague was using it for real work, and Cherny found himself working on it every weekend¹.

Until December 2024, development was essentially Cherny and Sid Bidasaria with a little help from Mann. After the green light, six or seven people jumped in for a final two-week sprint - and many of today’s core features, like bug reporting and the login flow, were built in those two weeks¹. In February 2025, the tool was renamed from Claude CLI to Claude Code and launched externally as a research preview¹. Head of product Cat Wu is candid about the early-access reception before launch: it was “lukewarm”¹.

What made it take off? Cherny points to two things arriving together: “a business model innovation coupled with a model innovation” - subscriptions plus the Claude 4 generation of models¹. Designer Meaghan Choi agrees: “When the Claude 4 models came out, that’s really when our moment came. … The model just wasn’t ready for the product that we wanted to build. But then it was”¹. The product mattered, but the base model crossing a capability threshold was decisive.

The Numbers: From 10% to 100%

The piece includes concrete numbers that capture how fast things changed. The share of Cherny’s own code written by Claude Code was still small around launch in February 2025, but climbed steadily - and by winter 2025: “100 percent of my code was written by Claude Code. Not a single line by hand”¹. Fortune reported in January 2026 that Cherny said the same on X - “it has been 100% for two+ months now, I don’t even make small edits by hand” - and that an Anthropic spokesperson put company-wide AI code generation at 70-90 percent, with roughly 90 percent for Claude Code itself⁴. Cherny also mentions writing 88 commits in a single day just before the interview¹³.

User behavior shifted too. Wu observes: “When we first launched, everyone was reading every single permission request Claude Code made. These days, a huge portion of our users just auto-accept everything. I think the transition shows that Claude has earned their trust”¹. Outside Anthropic, the story features early adopters like Austin Ray, who drove adoption across fintech company Ramp, and Bun founder Jarred Sumner, who says he “changed how we prioritized at Bun to make it easier to use with Claude Code”¹. Large-scale deployments like the Government of Alberta auditing over 400 million lines of code with Claude sit on the far end of that same curve.

The Takeaway: Ship at 30%, Bet on the Next Model

For readers building or adopting AI products, the most transferable idea in the piece is a different definition of “done.” Mann puts it plainly: “you have to build something that works 20 or 30 percent of the time now, so that when the next model comes out, it works 80 percent of the time. And that’s enough to get market traction”¹. Build the imperfect product ahead of the model that will make it work - a view echoed by research engineer Dawn Drain’s line that “Once you cross the model capability threshold, the form factor kind of reveals itself”¹. Anthropic itself now says most of its code is written by Claude, a loop where the tool accelerates its own development.

On team building, Cherny’s reflection is notable: “keeping the team small was actually very important to our success. It helped us be really creative with resources, and kept us from overengineering. And it forced us to use Claude more”¹. So is the fact that auto-updates and detailed user metrics were built in from the earliest days, letting the team ship fixes minutes after feedback arrived¹.

As for what’s next, Igor Kofman predicts the next abstraction level: “you’re not managing a bunch of Claudes - now you’re managing the Claudes’ manager”¹. And Kravec offers a forecast worth sitting with: “for most of 2026 and 2027, there’s going to be quite a lot happening in as little as three months. … That’s the disorienting part, and I’m not sure if anybody is ready for it”¹. The original is a long seven-part read, complete with a terminal-style reading mode - worth the time if Claude Code is part of your daily work¹.