The Making of Claude Code
Anthropic published the oral history of Claude Code — from a safety-research side project nicknamed clide to the tool that changed how Silicon Valley writes software. The interesting part isn't the launch. It's the four quiet years before it, and the shrug it got after.
Most product origin stories are told backwards. Once a thing wins, every early decision looks inevitable and every detour looks like strategy. Anthropic's new oral history of Claude Code resists that tidy narrative. It is a story about a company that was not supposed to build products at all, an internal tool almost nobody outside the building used, and a public launch that landed with a shrug.
The piece is told by the people who built it — researchers, engineers, product folks, and early users, 19 of them across research, core building, product, launch, and adoption. Strip out the nostalgia and what is left is a sharper lesson about timing, dogfooding, and building a surface before the model can carry it.
A safety lab that wasn't supposed to ship products
Anthropic was founded in 2021 when Daniela and Dario Amodei left OpenAI with about a dozen core researchers. The founding vision was narrow on purpose: a pure AI-safety company. The remit was to research AI, not to build and sell products.
That is the first irony of the story. The organization that explicitly did not want to be a product company ended up building the developer tool that every other lab is now racing to copy. Claude Code did not come from a product roadmap. It came from people trying to do their own work faster.
clide: the tool before the tool
Before it had a name anyone would recognize, it was an internal tool nicknamed clide — a rudimentary VS Code extension that circulated inside Anthropic. It was not a launch candidate. It was scaffolding that researchers and engineers used on their own repositories, day after day, because it made them faster.
This is the part worth sitting with. The tool earned its right to exist internally, through use, long before anyone tried to turn it into a product. Nobody had to be convinced by a deck. The value showed up in the daily work of the people building the models.
Boris Cherny and the two-week sprint
Claude Code lead Boris Cherny joined in September 2024 and started iterating fast. The core of the tool as people would first meet it came together in a two-week sprint that December. That kind of speed is only possible when the groundwork already exists — the years of internal use, the model underneath, and a clear idea of what the thing wanted to be.
The launch that landed with a shrug
Claude Code launched publicly in February 2025. The response was lukewarm.
That sentence deserves more attention than it usually gets. The tool that people would soon credit with reshaping how Silicon Valley writes software launched to a shrug. If you judged it on launch-day buzz, you would have written it off. The launch was not the story. What happened next was.
Claude 4 is what actually changed
Adoption exploded after the Claude 4 model release. The harness did not transform overnight — the intelligence flowing through it did. Claude Code was, in large part, the same surface it had been at launch. The model crossed a threshold, and suddenly the surface that was already sitting there could catch all of that new capability.
That is the whole shape of it: a tool built slightly ahead of the model, waiting, and then a model release that made the waiting worth it. A product like this has two launch dates — the day it ships, and the day a model makes it obvious. February 2025 was the first. Claude 4 was the second, and it is the one everyone remembers.
The timeline in one view
What builders should actually take from this
The nostalgia is nice. The lessons are more useful.
2. Dogfooding beats market research. clide earned its existence inside the building first. Real daily use by the people closest to the model is a stronger signal than any survey.
3. A lukewarm launch is not a failed launch. February 2025 was a shrug. Timing against model capability mattered far more than launch-day noise. Judge tools on their trajectory, not their debut.
4. "1% done" is the honest posture. The tool that moved an industry, and the lead says it is barely started. That is not false modesty — it is what it feels like to build on a capability curve that keeps bending up.
The uncomfortable takeaway
The most-copied developer tool of this cycle came out of a lab that did not want to build products, started as a throwaway extension, and launched to indifference. None of the usual product playbook predicted it. What predicted it was proximity to a fast-improving model and the patience to keep a rough internal tool alive until the model caught up to it.
Boris Cherny says they are 1% done. On the evidence of the last four years, the interesting question is not whether that is true — it is what the surface sitting quietly in the building right now will catch when the next model lands.
