My First Look at Claude Code

Not behind, but at the forefront: On feeling overwhelmed by AI progress, and why that means you're exactly where you need to be.

I know it’s only been a few weeks since I wrote about the joy I find in writing code, but I’ve heard too much about Claude Code and other coding agents to ignore any further.¹

There’s a new course at DeepLearning.AI taught by Elie Schoppik, head of technical education at Anthropic: Claude Code: A Highly Agentic Coding Assistant. The course is free, and only takes about 2 hours.

I finished this course over the holiday break and I was left sitting with a feeling I suspect some of you share: this peculiar mix of exhilaration and dread that comes from watching the ground shift beneath your feet in real-time.²

This course AI-pilled me

Let me be clear. I’m not new to AI coding tools. For years I’ve been using GitHub Copilot for autocomplete, Positron Assistant for data science work, and I regularly bounce between ChatGPT and Claude’s chat interfaces when I’m stuck on a problem. I thought I had a handle on this AI-assisted development thing.

Not quite.

Claude Code isn’t just an incrementally better chatbot or autocomplete. This thing can autonomously plan, execute, and improve code for extended periods with minimal human input. The wildest part of this course for me was watching the instructor use git worktrees to run multiple instances simultaneously, each working on different parts of a codebase while he coordinated the effort. Here’s a short clip:

Excerpt from the course, Claude Code: A Highly Agentic Coding Assistant (deeplearning.ai, built with Anthropic; instructor: Elie Schoppik), fair use here for purposes of commentary. Clip from the video Adding Multiple Features Simultaneously, at about the ~5 minute mark. All rights in the original content belong to the respective rights holder.

The course used Claude Code to explore and jazz up a RAG chatbot, refactor Jupyter notebooks into Streamlit dashboards, and to build web apps from Figma mockups.

Watching Elie Schoppik work through all this felt less like learning a new tool and more like watching a fundamentally different way of building software.

The Vertigo of Progress and Reframing

After watching this I had the feeling of being irrecoverably behind.

The course covered techniques I hadn’t considered, integrations I didn’t know existed, and workflows that seemed to require a mental model I hadn’t yet developed. And these capabilities? They’re from now. Not some theoretical future. Right now, available today. Which means that in six months, a year, the landscape will have shifted again, and the skills I’m scrambling to learn today might be table stakes tomorrow. The feeling of falling behind despite actively learning is exhausting.

I mentioned this feeling of falling behind on Bluesky, then Jase Gehring (UC Berkeley) responded to my initial thoughts with something that stuck with me:

“@stephenturner.us: You’re not behind, you’re at the forefront. No one knows how to use these tools. The current capabilities were non existent a year ago.”

Read that again and let it sink in. No one knows how to use these tools.

Not really. Not comprehensively. Because the tools themselves are too new. The best practices are still being discovered. The workflows are still being invented. The entire field is learning in real-time, together, through experimentation and shared experience.

Claude Code only came out in limited preview in February 2025, and generally available just last May. None of this existed a year ago. None of it!

Here’s another way to think about it: if you’re exploring these tools at your job or for fun, you’re not just playing catch-up. You’re pioneering. I appreciated the comment from Jase. This reframing isn’t just self-serving feel-good rhetoric. I think it’s strategically important for a few reasons.

The new skills you’re developing right now are foundational. The people learning to effectively prompt, coordinate, and collaborate with these agentic tools aren’t late to the party. They’re establishing the patterns that will define how software gets built in the coming years.

The knowledge gap is universal. Your colleagues aren’t secretly experts at this. The senior developers at Big Tech Company are figuring this out alongside you. The playing field is more level than it’s been in decades because everyone is learning simultaneously.

Early adopters shape the future. The feedback, use cases, and best practices being discovered by people learning these tools now will directly influence how they evolve. You’re not just learning to use a tool; you’re participating in defining what these tools become.

What Now?

I suppose I need to end this essay with a call to action, so here it is. Stop telling yourself you’re behind. Recognize that you’re early.

Spend two hours and take the course or others like it. Experiment with Claude Code or whichever agentic tool speaks to you. Break things. Try workflows that seem absurd. Share what you learn. I plan to do so here in future posts.

Learning in Public

The anxiety you feel about the pace of change? That’s not a sign you’re falling behind. It’s confirmation you’re paying attention. It’s evidence you’re engaged with the frontier of our field. We’re in a moment of profound transformation in how software gets built. The developers who thrive won’t be the ones who had a ten-year head start mastering these tools (because that’s impossible).³ They’ll be the ones who embraced the discomfort of not knowing, who experimented boldly, who learned in public, and who helped shape the practices that everyone else will eventually learn.

The forefront is an uncomfortable place to be. It’s supposed to be. But it’s also where the most interesting work happens.

Coda: Software Engineering vs. Data Science

One important caveat: Claude Code really shines in traditional software development contexts where you’re building features, refactoring code, fixing bugs, or implementing designs. These are tasks where you can specify an endpoint, and an agentic tool can autonomously work toward it. But data science work often doesn’t fit this pattern.

When I’m doing exploratory data analysis in Positron, I’m not building toward a known specification. I’m running a few lines, looking at the output, making a plot, seeing something unexpected, and then deciding what to explore next.

The next line of code literally cannot be written until I’ve inspected what the current block produced.

This is a different type of tool like Positron’s Databot shines.⁴

Databot is a purpose-built exploratory data analysis agent. It writes and executes its own code on the fly to help you understand your data. It doesn’t care very much about the code you’ve already written.

The agentic approach of Claude Code is transformative for software engineering tasks (which in my case can include things like writing R or python packages that help with some data science / data analysis task). But the data analysis side of data science requires a different kind of collaboration where the human stays tightly in the loop, making judgment calls based on what they’re seeing in real-time. It’s not that one approach is better; they’re solving fundamentally different problems. Understanding which tool fits which workflow is part of that pioneering work I mentioned earlier.

1. As I was wrapping up this essay, Ethan Mollick at One Useful Thing published a really nice piece on Claude Code and What Comes Next. I’d recommend reading this one. Also, a few hours before hitting publish on this essay, Matt Lubin wrote about Claude Code at the Bio-Security Stack, which I also recommend reading.

2. Turns out I share this feeling with Andrej Karpathy who recently posted: “I've never felt this much behind as a programmer. The profession is being dramatically refactored as the bits contributed by the programmer are increasingly sparse and between. I have a sense that I could be 10X more powerful if I just properly string together what has become available over the last ~year and a failure to claim the boost feels decidedly like skill issue.”

3. Although this won’t stop employers from asking. My favorite example of this was from an IBM job ad in 2020 requiring 12 years of Kubernetes experience, when Kubernetes was only 6 years old at the time!

4. I’ll write something here soon about Databot. But for now, consider the fair warnings from Positron that great power comes with great responsibility. And this choice quote from Joe Cheng, CTO at Posit: “In my 30-year career writing software professionally, Databot is both the most exciting software I’ve worked on, and also the most dangerous.”