Today, I used Claude Code, an experimental tool released by Anthropic a while ago. I wanted to try it right after the announcement, but I kept putting it off because I didn’t have a concrete use case. Today I had to rapidly prototype an app, so I finally gave it a shot.

Claude’s agentic capabilities

Claude Code is basically Claude 3.7 Sonnet wrapped with tools for command-line work. Claude can see your repository, view your file directory, edit files, and execute terminal commands. So it has more agency than a normal chat interface.

I needed to quickly prototype an app that hosts images, allows filtering by tags, and enables image similarity search. After just 2-3 hours, I had a functioning prototype.

I haven’t used Cursor, the VS Code variant that gained traction and was widely praised by people in the Bay Area. But maybe now I’m experiencing the epiphany that Cursor users had. The experience was not just typing a query and having Claude do everything from start to finish. It felt more like pair programming.

It gave me the sense that there was a person behind the terminal, troubleshooting and debugging with me. Claude wasn’t perfect, of course, but the back-and-forth was refreshing. It felt much better than randomly encountering errors and sifting through Stack Overflow links.

A new kind of collaboration?

It was a time-saver. I provided high-level directions, Claude executed them, and when problems arose, we solved them together. The biggest surprise was how collaborative the process felt. It was much easier than doing everything on my own or switching back and forth through a chat interface.

If I had done this single-handedly, it would have taken days. But for about $3.50 in API costs, which might seem expensive depending on your threshold, I built a prototype that I’m pretty confident in within 2-3 hours.

Claude Code also helped me fix bugs on my self-hosted blog. I’m more of an AI guy, not a web guy, so managing my blog often involved trial and error, throwing things at the wall and seeing what stuck. I had to manually fix bugs and do all sorts of tweaks to get it to look the way I wanted.

But with Claude Code, I could query my repository, and the debugging became conversational. That made it much easier to fix persistent issues quickly. All in all, it’s a useful tool.

Collaboration as a mirage

But the reason it felt collaborative is that Claude still isn’t perfect in its agentic execution. Don’t get me wrong: Claude is capable, but it still requires human intervention. For complex tasks, it doesn’t always get things right on the first try.

If we ever reach a point where AI no longer needs intervention, where it can figure out even highly complex instructions entirely on its own, then it would no longer feel like collaboration; it would be full automation.

That said, I structured my app development process in a very step-by-step manner, executing tasks sequentially. Maybe that’s why I didn’t run into any major setbacks. If I had given more abstract instructions, I might have been more surprised by what Claude could do. But the key point remains: it still feels like collaboration because Claude makes mistakes, and I had to intervene.

Looking at the trajectory of AI progress, though, I don’t think this collaborative phase will last long. Eventually, it won’t feel like a partnership. It will feel like replacement.

The inevitable shift

I’ve recently seen posts on X about recruiters at frontier labs assuming that junior staff are AI-replaceable. I was initially skeptical of this claim, but now I kind of believe it. Also, in Andrej Karpathy’s post, he argues that agency is more valuable than intelligence. This struck me as profound.

In terms of intelligence, today’s top LLMs are already more knowledgeable than humans. But in terms of agency, humans still have the upper hand. After using Claude Code, I caught a glimpse of a future where machines may reach agency supremacy as well. When that happens, it won’t be collaboration. It will be full Humans Need Not Apply territory, outdated but still relevant.

I encourage everyone reading this to try it for themselves. Hearing about it isn’t the same as experiencing it firsthand. Maybe I’m exaggerating, but I stand by my belief that this kind of automation is a matter of when, not if. This is the worst this technology will ever be.