Claude’s Anthropic Code Head Says Agile Engineering Isn’t That Important
YY Combinator Diana Hu recently spoke with Boris Cherny, creator and head of Claude Code at Anthropic. Cherny said that Claude Code is more skilled than is often realized, saying that rapid engineering is less important. Since Claude doesn’t need to “overthink,” he suggested that Claude Code users would be better off not following AI advocates on Twitter and LinkedIn.
Agile Engineering is Less Effective Today
One point that Cherny is happy about is that Claude Code is highly skilled and does not require careful information. He gave an example of an experiment that went viral in which a developer gave Claude access to OpenCV (Open Source Computer Vision Library), an open source library of several thousand real-time vision algorithms for image processing and machine learning. Then it was asked to draw pictures, and it did, a totally unexpected ability.
The point of the anecdote was to reinforce the message that Claude Code has skills that far exceed the need for “quick engineering.” Diana Hu asked about model extraction, which is the process of unlocking hidden capabilities in AI models.
Diana Hu asked:
“And the big area of research for this is basically invoking the model, right? Being really good at getting all these skills and asking the model to do the right thing, right?
How do people get better at that? And effectively, how do people get better at agile engineering? Do people still need to do more engineering quickly or is that changing too?
Tell us where this is going.”
One of the things Cherny kept repeating is that it helps to give Claude code tasks that are difficult to do and exceed your expectations of what Claude is capable of, which was the point of the OpenCV anecdote.
Answer from Cherny:
“Yes, I remember last year, one of the most popular jobs was the agile engineer. And then it changed and I thought it became like a content engineer. So there are these kind of waves of it, I think it’s like coming and going.
I guess skill these days is less about quick engineering and more about figuring out how to give Claude a seemingly overwhelming task? And how do you get Claude to validate his work along the way?
And validation, I think, is probably the single most important thing that people don’t get right, especially. “
At this point, Cherny described giving Claude Tag (Claude who works as an agent within Slack) access to a Mac virtual machine and gave it a complex programming task. He asked her to rewrite the Claude desktop app, converting it from the open-source Electron framework to Apple’s programming language called Swift. Cherny used simple, direct information that did not think too much about instructions or constraints.
Here is his job description and information:
“Then I’d say, OK, now what I want you to do is I want you to rewrite the Electron app in Swift. I want you to run the Electron app in a virtual Mac machine, capture it, and look pixel by pixel, compare it to the Swift version.
Don’t stop until you finish.
…So this is one of those examples where a model can do it today, you just have to let it do it, and you don’t need fancy things.
You don’t need /goal, you don’t need /loop. This helps. But really all you need to do is give the model a function, give it a way to validate the output of its function so it doesn’t get stuck, and it’ll just go.
…And actually, this time, Claude decided to live blog it. So what he did was create a Slack channel internally, and started posting screenshots every few minutes of his progress. “
How Claude Code Users Can Become Professionals
The interviewee, Diana Hu, expressed surprise that such a complex planning task could be accomplished so quickly. He then asked how users of the Anthropic Claude Code could become more proficient with the Claude Code. Cherny’s response was that the promotion of Claude Code should be done “by force,” meaning by experience and observing how things go.
He asked:
“What separates the people here who are Claude’s top 1% users? How can people learn to use Claude like Boris?”
Answer from Cherny:
“Maybe like, don’t listen to LinkedIn influencers. Don’t listen, don’t read Twitter.
This is the thing about modeling that I think everyone wants as one weird trick to do. That’s not there, there’s nothing like that.
The way the model works you have to approach it rigorously. You have to give very hard work.
You have to give it the tools to validate the work, just like you would do it yourself, just like you would if you were doing the work.
You have to see where it struggles, and then you have to fix that, with better motivation or skill, or if the model lacks context, give the MCP to be able to pull in the situation we need. It is what it is.”
Over-Specification Is Self-Defeat
The interviewee, Hu, commented that it sounded easy. Cherny responded that he thought “people tend to overthink.” His point was that over-specification is unnecessary with today’s Claude Code.
He continued:
“I think people tend to over-engineer. Because I think, in many ways, when we built systems in the past, that’s the way you had to do it.
So when I look at developers who have been coding for a long time, years or decades, this is a really common failure mode, trying to specify, and trying to overspecify, and get the model to do the job the way you could.
And that’s just not how the model works.
But I think a lot of people let this go, and it’s a journey to let go. It’s a journey of figuring out how to handle this thing like you would a co-worker.
I think that’s the level of intelligence that it’s at now.”
What is taken
- Fast engineering is more important than giving Claude clear instructions and methods to ensure their work.
- Claude has progressed so far that many “quick engineering” tricks are no longer needed.
- Code Claude is capable of more than most users realize.
- Challenge Code Claude for difficult tasks.
- Trying things out and seeing how they work can be helpful.
- Don’t think too much about the instructions.
Watch Boris Cherny’s Interview:
Featured image by Shutterstock/RixAiArt
