Boris Cherny at the Computer History Museum: 7 Lessons From Claude Code's Creator

Boris Cherny built Claude Code. In September he sat down at the Computer History Museum for 72 minutes, and barely 10,000 people have watched it. It's the most honest we've heard him, so we cut it down to the moments that matter.

The short version: writing code is getting cheap. Knowing what's worth building isn't. Here are the stories behind that line, in the order he told them, and what each one means for you even if you never write code.

His prediction: anyone codes like him in a year

The talk's boldest line comes about halfway through: "My prediction would be in a year, maybe less, everyone will be able to write code as well as I can." He says the gap between a professional who has studied software for a whole career and someone who has never written a line of code will keep shrinking, "so that anyone that wants to build software will be able to build software."

The kid who cheated with a calculator

Cherny started coding at about 13, on a TI-83 calculator in math class. His first programs were the answers to a math test, typed into the calculator. When the math got harder, he wrote small programs to solve the questions, then switched to assembly, a lower-level way of programming, "so I could cheat better on the math test." He even copied the programs to his classmates' calculators. Eventually the teacher caught them all.

Notice what that story says. The code was never the point. The result was. Hold on to that, because it's the whole story.

The bet: model overhang

After dropping out of college, building startups and working at Facebook, he joined a tiny team at Anthropic called the labs team. Its idea: "The model has capabilities that the product does not let the model express." They called this model overhang. In plain words, the AI could already do far more than any app allowed. Back then, AI coding meant pressing tab to finish one line.

Six people built Claude Code, MCP and skills

That team was "six people or eight people or something." Over a few months, it produced Claude Code, MCP (the standard way to connect Claude to other tools), the Claude desktop app, the first versions of computer use, and skills. That's most of what you use with Claude today.

Stuck at $40 million for months

You'd think it took off right away. It didn't. "This number is seared into my head," he says. Revenue sat at $40 million for months. "It was just flat, and we couldn't figure out how to move it." The team stayed up late every day for three months and worked every weekend.

What finally moved it wasn't a new feature. It was a smarter model. When Opus 4 and Sonnet 4 came out, growth turned upward "because people found it useful for the first time. The model would do the right thing." He now describes Claude Code as "definitely billions in revenue, millions of users."

What this means for you: if an AI tool let you down six months ago, the tool might not be the problem anymore. Try it again with the same task.

He stopped writing code by hand

One November, Cherny realized he hadn't edited code by hand in a couple of weeks. Claude had been writing all of it. "I just didn't even realize it."

His normal day now includes a data analysis session that can run 12 hours. Claude forms a guess, checks the data, rejects the guess, and "launches a fleet of 10 agents to investigate something in parallel," then checks their findings and throws some out. An agent here just means a copy of the AI working on its own task. And that, he says, is one of many running at once.

Claude messaged six employees on Slack

This is our favorite story in the talk. Cherny wanted to know who at Anthropic wrote code but wasn't using Claude Code. He asked Claude to check the company's data. It found six people. Then he asked it to message each of them on Slack and ask why.

Everyone replied within about 20 minutes, confused: they used Claude Code every day. Claude's response, as he tells it: "I think it's a logging bug." Then it fixed the bug.

He says average employees now have hundreds of agents running every day, and adds: "I don't think there's any human-written code left."

The errand anyone can hand off

Not a programmer? This one is for you. Cherny and his friends wanted to go clamming near Seattle and couldn't work out how to get a license from a confusing local city website. So he asked Cowork, Claude's tool for everyday tasks: "Can you get a license for me?" It ran for about 30 minutes, found the site, worked out how to use it, and got him the license.

No code at all. Cowork is now built right into the regular Claude app.

Why memory took a year and a half

Getting Claude to remember the right things took the team a year and a half, and the early failures are funny. While building a website, he told Claude one button would look better in blue. Claude's reply: "Great, the button should always be blue. I'll remember that." He meant that one button.

The yes button is security theater

You know the little prompt. Claude wants to run something and asks you, yes or no. It's there to keep you safe. Anthropic tested whether it actually does.

They hired contractors to solve coding puzzles with Claude Code, and now and then slipped in commands that would have caused damage if they were real. People "accepted it almost all the time." Cherny admits he does the same: "I just say yes, yes, yes, yes."

His conclusion: "This has become security theater. It's not real security, because no one is reading the commands anymore."

The fix is auto mode. It takes the thing Claude wants to do and asks a second model, one with none of your conversation's context, a single question: is this command safe or not? "It gets it right almost every time," he says, "compared to a person that gets it right almost none of the time." He also says you can simply ask Claude to change the rules: what's allowed and what isn't.

Prompt injection: a $20,000 prize, zero breaks

Prompt injection is when hidden text in a web page or file tricks an AI into following someone else's instructions. Cherny says that in Claude Code, Anthropic "can no longer demonstrate a successful prompt injection." They hired outside security researchers and offered a $20,000 prize to anyone who could find one in a week. Nobody did.

To be fair

He's careful here, and we should be too. His exact words: "I think the vulnerabilities probably still exist, but they're really hard to find now." Zero breaks in one week of testing is not the same as unbreakable.

The same goes for the rest. Every number in this post is what he said on stage, about his own product. The prediction that anyone will code as well as he can within a year is a bet, not a result.

The jetpack and the printing press

Asked how the work feels, he says: "This is the most fun I've ever had building. It just feels like I have a jetpack. Any idea I have, I can just build."

He backs the prediction with history. In the 1400s, he says, 2 or 3 percent of Europe could read. Reading and writing was a job, done by scribes working for lords and kings. Then came the printing press. In the next 50 years, Europe produced more literature than in the 1,000 years before, and the price of a book fell about 100 times. The press didn't end reading. It made everyone a reader. His bet is that code is next.

Why experienced engineers struggle

Cherny says he sometimes hires engineers with decades of experience who "make all the wrong decisions" because they're used to how things worked before. The ones who succeed need about six months to unlearn it. Meanwhile, some new graduates teach him things: "They trust the model more than I do."

What he hires for now

Two things, and neither is a coding skill:

• Curiosity. "I don't think it's the coding that matters that much anymore." What matters more is being curious about the field, the business and what the people around you are doing.
• Judgment. As the pace gets faster, nobody can manage you closely. He wants people who know the right thing to do and who say so when they're stuck.

What he would tell a CS student

The last question of the night came from the audience: what would you tell a second-year computer science student, keep studying or switch departments? His answer had two parts. First, be curious and use all the tools: "Don't shy away from agents." Second: "Probably get a minor in something else."

Remember the kid with the calculator? He never loved code. He loved what it got him. That's the whole message.

Get the Jetpack Playbook

We turned the talk into a free one-page playbook: all seven of his lessons, with the minute to jump to in the talk, and what to try this week even if you never write code. Comment JETPACK on the video, or grab it below. The full 72-minute talk is linked here too.

Source: "CHM Live | Anthropic's Boris Cherny, Creator of Claude Code", Computer History Museum, moderated by David C. Brock. Quotes are from the public talk and lightly trimmed for reading; numbers are as he stated them on stage. Independent, not affiliated with Anthropic or the Computer History Museum.