← Journal

Writing Code with LLMs

Lately I’ve been messing around with Claude Code to see what it’s actually like building something with an LLM.

I’ve used Claude Code, Loveable, and other tools for prototyping ideas. That’s fine and useful. But I really wanted to push the limits of going from idea to production. What would the process be like? How would I feel about it? And how much would I actually understand about what’s happening under the hood?

So I decided to try spinning up a small marketing site and play around with a couple of frameworks:

  • Next.js
  • Astro

Just to see how far I could get with prompts.

From nothing to deployed

I asked Claude to create the project and get it running: set up the repo, add the scaffolding, get it running locally, and configure deployment.

Here's what it ended up doing:

  • Created a new GitHub repo
  • Scaffolded a Next.js site
  • Set up components and SCSS
  • Installed dependencies
  • Ran everything locally on localhost
  • Pushed the repo to GitHub
  • Deployed it to Vercel
  • Created staging URLs

All from a handful of prompts.

The whole thing took about 45 minutes.

Most of that time was just waiting while it did its thing.

Normally this is the part that burns a day — fiddling with configs, environments, deployment setups, random errors. Instead it just… worked.

The part I thought I'd hate

Before trying this I had one worry.

I thought I'd hate building something without really knowing what was going on under the hood.

But that hasn't been the case at all.

I still understand the whole thing. The project structure, the components, how the site runs locally, how the deployment works.

The difference is I didn't have to deal with all the annoying setup stuff:

  • Local environments
  • Dependency issues
  • Deployment configs
  • Project scaffolding

That stuff doesn't make you better at building things. It just slows you down.

What actually changed

The interesting part is what you end up spending time on instead.

Once the setup friction disappears, you can just… build.

I found myself thinking about the structure of the site, refining the content, trying different layouts, and testing random ideas just to see what happens.

The feedback loop becomes ridiculously fast.

Think of something → prompt it → see it running a minute later. Is there even a need to write tickets for work if you can just think and execute? (Joking, of course, but you get the idea.)

The takeaway

The best way I can describe this:

It feels like learning to speak a new language, except you're fluent in five minutes

You still need ideas. You still need taste. You still need to know what good looks like.

But the distance between thinking something and building it just collapsed.

Which honestly feels kind of crazy.