Claude Code Is Great. It Wasn't Right for Me.

Everyone is talking about Claude Code right now. If you haven’t heard of it, it’s Anthropic’s command-line coding tool — you describe what you want, and it builds it. Autonomously. It writes files, runs commands, debugs errors, and delivers a working project. For developers, it’s a game-changer. For someone like me — a non-developer with big ideas and limited coding knowledge — it sounded like a dream.

And it was, briefly. Then I realized it was the wrong kind of dream.

The CRM Experiment

A while back, I had an idea for a CRM. I’ve worked with customer relationship management systems for over twenty years in legal tech. I know exactly what’s wrong with most of them, and I’ve got strong opinions about how to fix those problems. What I don’t have is a computer science degree.

So I turned to AI. I wrote a few paragraphs describing what I wanted — the features, the workflow, the pain points I wanted to solve — and asked Claude Code to build it. And it did. It generated files, set up a database structure, created the interface. It was impressive.

But when I looked at what it produced, I realized something uncomfortable: I had no idea what any of it meant. I couldn’t have told you why one file existed versus another, what the database schema was doing, or how to change something if I wanted to tweak it later. I’d essentially asked someone to build me a house and then stood outside admiring it without knowing where the plumbing was.

The CRM project fizzled. Not because the code was bad — it wasn’t — but because I had no relationship with it. It was someone else’s work that happened to match my description. I couldn’t maintain it, extend it, or learn from it because I’d skipped the entire process of understanding it.

The Slow Way

Fast forward to this week. I’ve been building a website for a community project I run in Second Life — a virtual world I’ve been part of for years. The site is built on Laravel with a tool called Filament, and I’m constructing it piece by piece in conversation with Claude.

And by “piece by piece,” I mean it. There is a lot of copying and pasting. I describe what I want a page to do. Claude explains the approach, writes a block of code, and tells me where to put it. I copy it into my editor. I run it. Something breaks. I paste the error back. We talk about why it broke. Claude explains what went wrong, I learn something, and we try again.

It is, objectively, slower. It’s probably a little painful for both of us. There are moments where I’m sure Claude is thinking “I could have done this whole thing in thirty seconds if you’d just let me.”

But here’s the thing: I’m learning.

When we added a user management panel, I understood what a “resource” was in Filament and why it was structured the way it was. When we set up the navigation, I learned about route grouping and menu builders. When something broke because of a missing migration, I actually understood what a migration does — because we’d talked about it, not because Claude had silently run one in the background.

By the end of the session, I didn’t just have a website. I had a website I understood. I could look at a file and know what it was for. I could anticipate what might need to change next. I wasn’t standing outside admiring someone else’s house — I was inside it, and I knew where the plumbing was.

The Point Isn’t That Claude Code Is Bad

Let me be very clear: Claude Code is a remarkable tool. For professional developers who already understand architecture and just need to move faster, it’s transformative. For rapid prototyping, for experienced engineers who can read and evaluate what it produces, for teams that need to ship quickly — it makes perfect sense.

But for someone like me, whose goal isn’t just to have a finished product but to understand what I’ve built, the collaborative approach — the slow way, the back-and-forth, the copy-paste-break-fix-learn cycle — was infinitely more valuable.

AI isn’t just a tool for getting things done. It’s a tool for learning. And how you choose to interact with it determines which of those things you get.

Two Modes of AI

I think there are really two modes of working with AI, and most of the conversation right now is focused on only one of them:

Mode 1: Delegation. You hand off a task. AI executes it. You get a result. This is Claude Code, this is autonomous agents, this is the future everyone’s excited about. And it’s genuinely useful.

Mode 2: Collaboration. You work through a problem together. AI explains, suggests, teaches. You do the work, make the mistakes, ask the questions. You come out the other side not just with a deliverable, but with knowledge you didn’t have before.

Both are valid. But they produce very different outcomes. Delegation gets you a product. Collaboration gets you a product and an education.

For me, with the CRM, delegation felt hollow. With the website, collaboration felt like the whole point.

Try the Slow Way

If you’re someone who’s been using AI primarily in delegation mode — writing a prompt, getting an output, moving on — I’d encourage you to try slowing down. Ask Claude to explain why it’s suggesting something, not just what to do. Copy the code yourself instead of letting it run automatically. When something breaks, sit with the error for a moment before asking for a fix.

You might be surprised how much more you get out of it.

It won’t be as fast. It won’t be as smooth. But you’ll come away from the conversation knowing something you didn’t know before. And that, to me, is the real promise of AI — not just that it can do things for us, but that it can help us become people who can do more things for ourselves.

One prompt at a time.

— Drew