Claude Code Developer Boris Cherny: 'Build for the model of 6 months from now, not today's model'
"I thought this could really be something. I didn't know if it could write code yet, though."
Boris Cherny, the creator of Claude Code, felt that way in September 2024, working through weekends and every night for three months without taking a single day off. At the time, the model could only write about 10% of his code. Even so, he bet on the "model of six months from now." That bet paid off; today, Claude Code accounts for 4% of all public commits, and within Anthropic, it has increased productivity per engineer by 150%.
1. Claude Code started by "accident"—it was just an API experiment at first
1-1. The terminal was supposed to be a "starting point," not an "end point"
Boris Cherny stated at the beginning of the interview:
"What's unbelievable is that we are still using the terminal. That was supposed to be the starting point; I didn't expect it to become the end point."
The development of Claude Code was an extremely accidental process. Anthropic has long held the vision of "reaching safe AGI through coding," and the "Anthropic Labs" team that Cherny joined created three products:Claude Code, MCP (Model Context Protocol), and a desktop app.
However, at first, no one told him to "build a CLI tool."
"I just felt like the models were ready to build coding products. But no one had built a product that leveraged this capability yet. So, I started hacking. What was the first thing to do? It was to understand how to use the Anthropic API. I had never used it before."
Cherny built a small terminal app, for the simple reason that he didn't need to build a UI. It was a chat app. And when the Tool Use feature was introduced, he tried giving the model a bash tool.
"I asked the model, 'What music are you listening to?' It wrote an AppleScript to script the Mac and retrieved the music player information. I didn't think the model could do that. That was my first experience of 'a glimpse of AGI.' I realized the model just wants to use tools."
This happened in September 2024, in the era of Sonnet 3.5.
2. Latent Demand—everything from Claude.md to Plan Mode was taught by users
2-1. "The most important principle in a product"
Cherny repeatedly emphasized the concept of "latent demand."
"People only do what they are already doing. You can't make them do new things. Making what people are already trying to do easier—that's a good idea."
All of Claude Code's features were born based on this principle.
Claude.md was born because users were writing markdown files themselves and having the model read them. Cherny saw that and implemented it as an official feature.
Plan Mode is the same.
"I was looking at GitHub issues and internal Slack feedback channels at 10 PM on a Sunday. I saw users saying, 'Think of an idea, make a plan, but don't write the code yet.' I wrote Plan Mode in 30 minutes and shipped it that night. It was released by Monday morning."
The implementation of Plan Mode is surprisingly simple.
"There is no great secret to Plan Mode. You just add a sentence to the prompt that says 'don't write code.' In fact, you can even say that yourself."
2-2. "Plan Mode might become unnecessary in a month"
When the interviewer asked, "Won't there be no need to explicitly prompt for Plan Mode in six months?", Cherny replied as follows.
"It might be in a month. Plan Mode will probably end with Opus 4.5."
According to him, Claude Code is already becoming capable of entering Plan Mode on its own, and they are currently adjusting it to do so at the same time a human would want to enter Plan Mode.
3. "Don't build for today's models, build for the models of six months from now"
3-1. Scaffolding will become meaningless with the next model
What Cherny repeatedly emphasized was the principle of "don't optimize for the current model."
"The mindset at Anthropic is not to build for today's models. It's to build for the models of six months from now. This is my advice to founders who are still building on LLMs today. Think about the frontiers that models aren't good at yet. Because the models will become good at them."
He explained the concept of "scaffolding." You can write code around a model to improve performance by 10-20%. However, when the next model comes out, that improvement disappears.
"You can build scaffolding to get a performance boost and then rebuild it—or you can wait for the next model and get it for free. Claude.md and its scaffolding are examples of that. That's why I stuck to the CLI, because I didn't think I could build a UI that would be relevant in six months. The model improvements were just too fast."
3-2. The lifespan of a codebase is 'a few months'
"There isn't a single part of Claude Code that existed six months ago. It's constantly being rewritten. Every few weeks, we unship (remove) tools and add new ones."
When the interviewer asked, "Is 80% of the codebase from within the last few months?", Cherny replied, "Yes. Probably within a few months."
This indicates the "shelf life of code" in software development in the AI era.
4. Productivity at Anthropic—70% growth, and then 150% growth
4-1. 70% increase in PRs per engineer
Cherny showed the productivity improvements within Anthropic with concrete numbers.
"The team doubled last year, but productivity per engineer grew by about 70%. We measure this by the number of pull requests. We also cross-checked it with the number of commits and the lifespan of commits. Since the arrival of Claude Code, productivity per engineer at Anthropic has grown by 150%."
During his time at Meta, he was responsible for the code quality of all products (Facebook, Instagram, WhatsApp, etc.) as the lead for code quality. At the time, it took hundreds of people a year to improve productivity by 2%.
"The 150% figure is an absolutely unprecedented level."
4-2. "I uninstalled my IDE"
Cherny himself has completely deleted his IDE since Opus 4.5.
"I am 100% Claude Code and Opus only. I don't edit a single line of code by hand. I submit 20 PRs every day. Across all of Anthropic, 70-90% is written with Claude Code. For some teams, it's 100%."
In May 2024, he surprised his audience by predicting, "You won't need an IDE to write code anymore." However, for him, it was simply an extension of the exponential curve.
5. 2026 Prediction—The title of "Software Engineer" will disappear
5-1. Coding is "effectively solved"
Cherny shared his predictions for 2026.
"Coding will be generally solved for everyone. Today, for me, coding is effectively solved. It will be that way for everyone, regardless of the domain. I think the title of 'Software Engineer' will start to disappear. We might become 'builders' or 'product managers,' or the title might remain as a vestige, but the work people do will no longer be just coding. Software engineers will write specifications and talk to users."
This is already happening on his team. PMs, designers, EMs (Engineering Managers), and even finance staff all write code.
5-2. The upper-bound scenario—ASL4 and recursive self-improvement
On the other hand, Cherny also touched on the "upper-bound scenario."
"It is possible to reach ASL4. This is the stage where the model recursively self-improves. If this happens, there are standards that must be met before release. In extreme cases, catastrophic misuse could occur, such as the design of biological viruses or zero-day exploits. We are actively working to ensure that doesn't happen."
Conclusion—4% of all commits born from "coincidence"
Boris Cherny started with the "cheapest option," the terminal, kept observing latent demand, and kept betting on the model of 6 months from now. As a result, Claude Code accounts for 4% of all public commits and has become a tool used even for designing the course of NASA's Perseverance Mars rover.
What he repeated was the message, "Bet on the model." Bet on improving the model's capabilities, not on scaffolding. Believe in the performance of 6 months from now, not today's performance. And thoroughly observe what users are already trying to do.
"I just wanted to build something cool, and it was a surprise and very exciting that it became something truly useful."