It’s been three days since I started using Claude Code. The initial excitement was strong enough that I immediately wrote a blog post about it. After a couple more days, I have more thoughts, so here’s a quick follow-up.

Time-saving

I know I’ve already emphasized how much time Claude Code saves. Over the past couple of days, I’ve been completely revamping my personal website from the ground up. Let me preface this by saying: I’m not really a web guy. Before Claude Code, my blog was just a clone of the Al Folio theme. Whenever I wanted to tweak something, it felt daunting. Each tiny modification usually ended in frustration or compromise because the effort required to change something small just wasn’t worth it.

Now, if I want to adjust the tag styling or add new functionality, I just instruct Claude Code to do it. It handles the rest. The difference in my blog over just these past couple of days is significant. I’ve added small features and design tweaks that previously felt too annoying to attempt.

Of course, Claude isn’t perfect, but the amount of time it saves is substantial. Yesterday I wasn’t satisfied with how the tag feature functioned, so I decided to change it up. Claude handled the entire process for about $1 in API costs. A dollar. Less than a cup of coffee. Without it, the same change would have probably taken me hours of tinkering and frustration.

At first, Claude felt like a pair-programming partner, but even after just three days, that feeling is fading. It has become more like a virtual assistant, except this assistant can actually do things, not just talk about them.

Chatbots vs. agents: “what is” vs. “do it”

Using traditional chatbots like ChatGPT always felt like carefully constructing queries and specifying every little detail of my intent. Most of my interactions with ChatGPT, and likely yours too, are fundamentally knowledge retrieval: asking “What is this?” or “Tell me about that.” ChatGPT compresses a lot of internet knowledge into succinct answers. It is excellent at answering questions and explaining concepts, but it is mostly passive.

Claude Code changed that for me. With a truly agentic model, I don’t have to spell out exactly how I want something done. I specify what I want done, and Claude figures out the how. It summarizes information, but it also plans, makes decisions, and implements changes. The shift from “what is” to “do it” feels like the real jump.

Of course, because Claude isn’t perfect yet, the collaboration still exists in the form of verification. But as I’ve said before many times: verification is easier than generation, and right now, I’m comfortably on the verification side. I check Claude’s implementation by deploying my blog and testing the functionality. This still requires input, but the mental load is much lighter.

From collaboration to automation

While it still feels somewhat collaborative now, I’m realizing that my role is already shrinking from active collaborator to passive verifier. It’s an assistant relationship, not really a partnership anymore. Verification feels simpler and simpler, and I can foresee even verification becoming automated soon. At that point, human involvement becomes negligible.

Previously, I believed in the paradigm that self-feedback loops, generation versus verification, would incrementally increase intelligence. But intelligence and agency are different. Learning how to be smart or knowledgeable and learning how to act are not the same. The latter may be far more powerful. I’m now convinced that agentic capability will matter faster than I initially expected.

If, or more realistically when, verification itself is automated, we won’t be collaborating with AI. We’ll be observing it collaborating with itself. That’s when things get interesting and potentially unsettling.

Toward agentic intelligence?

OpenAI previously described levels of AI intelligence: Level 2 (reasoners), Level 3 (agents), and Level 4 (collaborators or organizations). Claude Code sits loosely at Level 3, and my recent experience makes Level 4, agents autonomously collaborating with each other, feel closer than I expected.

As these agentic models improve and increasingly handle their own verification, we’ll approach a point where human oversight matters less and less. Using Claude Code made that feel concrete.

As impressive as Claude Code is today, it’s the worst this technology will ever be. That is both thrilling and deeply concerning.