On July 6, 2026, Anthropic published a longform feature, "The Making of Claude Code." It is not written as a press release but as a seven-part oral history, told by co-founder Ben Mann, Claude Code head Boris Cherny, the development team, and early users, with interviews recorded between February and May 2026. The most counterintuitive part of the story is where it starts: the product that today writes the overwhelming majority of Anthropic's code began as a demo that collected "two or three likes" on an internal Slack channel. An internal demo that almost nobody applauded wrote its author's entire codebase a year and a half later — what this oral history really documents is a philosophy of products born ahead of their time.
The oral history reaches further back. The first product Anthropic ever built was not Claude but a coding assistant: a VS Code extension from spring 2022 that returned four suggestions per prompt and had roughly 100 external users — then the team shipped the API and "basically forgot about the coding assistant for a while." The research bet was placed even earlier. As head of reinforcement learning Shauna Kravec puts it, "we thought that the path to transformative AI would route through the ability to automate large chunks of software engineering work."
Between 2022 and 2023, the two pieces that make agentic coding work — a bash tool and search — came together, and an internal CLI tool called clide was born. In the interviewees' memory it was "wonky… very, very ahead of its time." When it first performed an autonomous edit, Adam Wolff, the first manager on the Claude Code team, was "dancing around my kitchen. I couldn't believe it."
The turning point came in 2024. In January, Ben Mann started the Labs team, seeing "a hole in the market for agentic coding." When Boris Cherny joined that September, he wanted to build something small — a linter — and Mann's reply is preserved verbatim: "No, no, no, no, you have to do the big thing." Boris pasted a rejected pull request into clide and watched it write the whole thing: "I had never seen anything like that. It felt like the future."
In about two days he built a demo he called Claude CLI and posted it on Slack. "I think I got two or three likes." But the very next day a teammate was using it for real work, and Boris found himself working on it every weekend. Until December 2024 the project was essentially him and Sid Bidasaria; after the green light, six or seven people ran a final two-week sprint — core features like bug reporting and the login flow were built in those two weeks, and auto-updates let the team ship fixes "five minutes" after feedback arrived. In February 2025 the tool was renamed Claude Code and launched as a research preview. Head of product Cat Wu describes the pre-launch early-access reception bluntly: "lukewarm."
Takeoff took two things arriving together. In May 2025 the Claude 4 generation landed, and designer Meaghan Choi recalls "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." Boris's summary is cooler: "a business model innovation coupled with a model innovation" — subscriptions plus a new generation of models. Outside evidence is equally strong: Austin Ray of Ramp concluded within the first five minutes that "this is going to fundamentally change everything," and Bun founder Jarred Sumner says he "changed how we prioritized at Bun to make it easier to use with Claude Code."
Around launch, only about a tenth of Boris's own code came from Claude Code. By winter 2025 his words were: "100 percent of my code was written by Claude Code. Not a single line by hand." Fortune reported in January 2026 that an Anthropic spokesperson put company-wide AI code generation at 70–90 percent, with roughly 90 percent from Claude Code itself; the day before this interview Boris had written 88 commits in a single day.
The shift in user behavior deserves the most attention. Cat 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." On the research side, Kravec already runs "a whole swarm of twelve different Claudes."
Judgment one: the best products are born ahead of the model. The most transferable line in the piece belongs to Ben Mann: "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." Research engineer Dawn Drain's version is more compact: "Once you cross the model capability threshold, the form factor kind of reveals itself." My reading: this is not a bet on luck — it is design that treats "models will keep improving" as the single most certain variable in the stack. The window belongs to early entrants; those who wait for perfect models are forever building for the previous generation.
Judgment two: a small team and eating your own dog food are the same thing. Boris's reflection deserves to be preserved as-is: "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." Auto-updates and detailed user metrics built in from day one compressed "feedback to fix" to five minutes. A small team is not a cost-cutting compromise but an accelerator of the product loop: the fewer and closer the users, the shorter the loop. Most companies treat scale as a moat; in this story scale is the enemy.
Judgment three: the real inflection point is not the automation of code but the automation of trust. The transition Cat Wu describes — from reading every permission request to auto-accepting everything — is at once the product's success and a relocation of its risk. The moment review is handed over, the safety boundary moves from "each concrete action" to "whether the designers' assumptions still hold." Eighty-eight commits in a day, a swarm of twelve Claudes, and Igor Kofman's predicted next abstraction ("you're not managing a bunch of Claudes — now you're managing the Claudes' manager") all amplify this line. Trust is earned through use; blind trust is trained by convenience — and on the interface, the two look nearly identical.
Reposting the oral history, Boris added only one line: "We are 1% done." In the interview he places Claude Code on a much longer lineage: "This is the IBM 029 — it's similar to what my grandpa programmed with in the Soviet Union. And this is one of the first text editors, which is still installed on every single Mac. And then it evolved, it evolved, it evolved… and somewhere on the spectrum is Claude Code."
As for what comes next, Kravec's forecast makes a fitting final frame: "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."
For anyone building AI products, the most honest lesson of this oral history hides in its origin: two or three likes, a two-day demo, a lukewarm reception — great products are rarely planned into existence. They are born early, and they live long enough to meet the model that makes them work.