Anthropic's First-Ever Look Inside Claude Code: Boris Cherny's Two-Day Demo Now Writes 100% of His Code

Before there was a product, there was a bet

Before Anthropic shipped any product at all, it had already bet on getting AI to write code by itself. The first thing they built was a coding assistant: a VS Code extension that took a prompt and gave you four different suggestions. By spring 2022, it had about 100 external users. The models were bad back then, and the infrastructure was primitive.

Shauna Kravec says the infrastructure for agentic coding (AI writing and running its own code) is far more complex than building a chatbot. Once code execution is involved, you have to figure out exactly what environment the code runs in and how to contain it safely and effectively. Her point cuts to the continuity here: many of the problems people run into with agents in 2026 are the exact same problems the team ran into in 2022.

clide, the internal tool that was both cool and painful to use

Before the real product existed, there was an internal tool called clide. The name was coined by colleague Eli Tran-Johnson for an even earlier tool. It was a command-line tool where you could chat with Claude and have it edit code and do dev work. Dawn Drain spent a long time teaching the model to write diffs (a representation of a code change — the most natural way programmers express edits), and clide grew out of that effort.

clide had a feature that let it fan out 100 Claude Haiku instances in parallel, so you could ask a question about an entire large folder even if that folder couldn't possibly fit in the context window (the amount of content a model can take in at once). Dawn Drain says he loved those pair-programming moments — pulling up clide and just answering a question on the spot, with people constantly asking him how he knew about all these useful tools.

Before moving to Labs, Adam Wolff added one last feature to clide, called clide agent. clide didn't have a bash tool yet at that point, so it was limited in what it could do. He got it to infer what you were trying to do from something you'd half-finished editing. It was very early-stage agentic behavior — the seed of something that could take the next step on its own. The day it first worked, he danced around his kitchen, barely able to believe it.

But everyone complained clide was clunky and slow to start (Sid Bidasaria). Boris Cherny says you had to recite a bunch of incantations just to get it running. It wasn't good software, but it had something magical and powerful about it — it had glimpsed the future. All the pieces were there; they just hadn't been put together yet.

A demo nobody understood, built in two days, changed everything a year later

Boris Cherny's starting assignment was four words: "automate programming." Nobody told him specifically what to build. He started by learning to use the API, tinkering as he went, and eventually put together a demo he called Claude CLI (command-line interface).

Nobody understood the demo — not even he fully understood it himself. But looking back now, all the original elements are still there. He asked it what music he was listening to, and it just screenshotted Apple Music and read it off — and that actually worked. This was roughly two days of effort. Redoing it today with Claude Code would take two minutes. He posted it to Slack and got two or three reactions.

The next day he walked into the office and saw Robert Boyce working, with those now-iconic red-and-green code diff lines up on his screen. Robert said: "Yeah, it's helping me write code." That was the moment it became useful. Ben Mann immediately realized: all the pieces existed, they just needed to be put together. After that, Boris started working on it every single weekend; friends would invite him out, but he couldn't stop thinking about it.

Sid Bidasaria says there were no pre-commit review gates on the codebase at all — if someone said "this doesn't work well," they'd push a fix within minutes and users would have it five minutes later. Boris had already built solid auto-update and good user analytics early on, and that's what made the whole loop actually spin. As a command-line tool rather than a web app, there was no complex architecture to navigate around — a simple client meant faster iteration.

Adam Wolff kept pushing back against growing the team too fast, while Boris kept wanting to expand it. Boris later admitted that, in hindsight, keeping the team small was actually key to the success: it forced everyone to conserve resources, avoid over-engineering, and lean more on Claude itself — otherwise there was no way to move fast enough. By the time of the official public launch in February 2025, Claude CLI was renamed Claude Code — a name suggested by product marketing's Alex Isken, and the team liked how simple it was.

The launch also carried two small hidden touches in the terminal. Late one night before launch, Igor Kofman thought it'd be nice to have an ASCII art logo, so he and Claude worked together filling out several character-art fonts, which became the now-iconic all-caps Claude Code logo that pops up on login — a little surprise at sign-in. Meaghan Choi's proudest touch was slipping a little character called Clawd into the terminal — a figure originally created by Sam McAllister for the Claude 3.5 Sonnet launch. She says chances to do something like that in a product don't come around often.

Ship something that can only pull off 20-30% today, so the next model generation can actually deliver

During the early access phase before launch, feedback was lukewarm. Cat Wu says people thought the idea was cool but there were bugs everywhere. They decided to ship it publicly anyway. Behind that decision was a counterintuitive product philosophy.

Meaghan Choi says the Claude 4 family was really the moment that belonged to them. Before that, the UX design they could actually do was limited — the model couldn't yet support the product they wanted to build; once it could, design finally had room to work. Boris Cherny adds the other half: they also launched a subscription model. Two engines firing at once — a subscription business model, and a leap in model capability — together ignited Claude Code's takeoff.

Dawn Drain puts it bluntly: Claude Code doesn't really owe as much to clide as people think. Once you cross the threshold of model capability, this form just emerges on its own.

Three people from three different worlds, and how they first started using it

After launching as a research preview in February 2025, it landed in three very different settings: desk-by-desk evangelizing inside a big company, an open-source author reworking his workflow around it, and a nonprofit getting access to custom software it could never have afforded before.

Tristan Hume's experience was the flip side. What he was working on at the time mostly required a lot of background knowledge — writing low-level kernels for accelerators, with a lot of detail that simply isn't documented publicly anywhere online. Early Claude Code wasn't great at building its own tools or learning on the fly, so for a while it wasn't much use to him — it could only handle very limited tasks. That only turned around once the models got stronger later on.

Now, nobody reviews permission requests one by one anymore

Just how much daily work habits have been reshaped shows up in a few concrete numbers. The share of Boris Cherny's own code produced by Claude Code kept climbing.

He remembers the day Sonnet 4 came out — he was backstage at the Code with Claude developer conference writing code, and felt the model had gotten so much better, so much more capable of acting on its own. By winter 2025, 100% of his code was written by Claude Code — not a single line by hand. He mentions one day he made 88 code commits, with his wife and dog by his side on the couch the whole day.

Shauna Kravec now runs an entire fleet of 12 Claudes at once, reading docs, updating things, pulling messages from Slack. As head of research, she hadn't actually written much code herself in recent years — now she writes more, because the barrier to entry is lower. Igor Kofman takes this a layer further: as the models and Claude Code keep improving, the next layer of abstraction is that you stop managing a bunch of Claudes and start managing the managers of those Claudes.

Cat Wu offers an observation about trust: at first, people would read every single permission request Claude Code sent one by one; now a large share of users just auto-approve everything. That shift shows Claude has earned their trust.

React shows us that anything successful eventually becomes something else

Wrapping up, several core people offered judgments based on their own experience — all comparisons grounded in fact, without mythologizing or forecasting.

Adam Wolff uses React as an analogy. React started as a very pure computer science idea — using functional programming to represent UI state, better than message passing. By the time it reached millions of daily active users, it had become something else: a logo, a brand, a feeling, far beyond the original technical idea. Most React users' reasons for liking it can no longer even be traced back to that original insight. His view is that Claude Code will evolve the same way: whether you think Claude Code is the terminal, Claude's personality, or some prompting technique — at the limit, none of that matters.

Tristan Hume pushed it to the limit once: he had Claude build a Jupyter Notebook replacement from scratch, without looking at a single line of code the whole time, even having it test the UI itself with a browser. It actually produced something usable. But when he used it himself, he found he didn't like it. He says he's waiting for a Claude with taste — one that knows everything he needs and can just handle it directly in the background.

Shauna Kravec's take draws on her background in theoretical physics. Some hard problems have taken humanity hundreds of years to work on, with progress far slower, often bottlenecked by expensive, hard real-world constraints — there's only so big a collider you can build. Her view: through most of 2026 and 2027, a lot is likely to happen within any given three months — the progress of a three-month stretch in 2024 was real too, but nowhere near this intense. She's not sure anyone is ready for that kind of mismatch.

Boris Cherny closes with a set of physical objects. He shows an IBM 029, much like the one his grandfather used in the Soviet Union; then one of the earliest text editors, still installed on every Mac today. It kept evolving, evolving, evolving, always changing. Claude Code sits somewhere on that same lineage.