← Journal

I Let an AI Agent Build Our Marketing Site for a Week

This week we ran a bit of an experiment at Shortcut.

I wanted to see what would happen if I mostly directed Claude to build a new marketing site for our company. Not in a vague way, but properly: giving it structure, telling it what to build next, reviewing outputs, iterating quickly.

The goal wasn't perfection. It was: how much can I realistically ship in a week if I stay in "director mode"?

The Setup: Changing How I Work

Normally, when I build something like this, I'm deep in it, writing components, adjusting styles, thinking about responsiveness line by line. Sometimes slow, but always controlled.

This time, I tried to stay out of that layer as much as possible. I focused on the structure, the product story, the flow, what pages needed to exist, and how everything should feel. Then I prompted Claude to build it, reviewed the output, tweaked it, and moved on.

At first, Claude wasn't great at the UI details. It was incredibly fast at a lot of other things, but UI wasn't its strength. In the beginning, I held its hand and directed when it should use certain spacing, sizing, colors, and other tokens. It felt less like coding and more like working with a very fast, slightly unpredictable junior engineer.

Monday — Building the Entire Foundation (Way Too Fast)

Monday was basically "set up everything", and it all happened in one stretch:

  • GitHub repo setup
  • Astro framework
  • Page structure
  • Vercel deployments
  • Header + footer
  • Global styles and themes (with keyboard shortcuts + toggle)
  • Layout system
  • Blog content imports

There were still real issues: font flashing, fixing fonts by making them local, things not working perfectly at first. So it wasn't magic. But none of it slowed me down for long.

Normally, one of these problems can derail an hour or two. Here, I'd describe the issue, get a fix, and keep moving. By the end of the day, the entire base of the site existed, which felt slightly ridiculous.

Tuesday — From Structure to Real Pages

Tuesday is where it started to feel like an actual product.

  • Connectors page
  • Security page
  • Footer refinements
  • Global styles improvements
  • Responsive testing
  • Kitchen sink page for components
  • Blog layout + styling
  • Pricing page

But what stood out wasn't just the output. It was how I was working.

Stream of consciousness building

Normally, I try to stay disciplined: finish one thing, then move to the next. This time, I didn't. It became fluid.

Mid-task, I'd think: "The keyboard focus states on dropdowns are probably off." Fix it. "This styling feels slightly off." Fix it. "I should build a proper test page for components." Build it.

No backlog. No "later". No pretending I'll come back to it.

Does this slow things down? If your goal is finishing one task as fast as possible, yes. But that's the wrong goal.

What actually happens is you solve X, Y, and Z at the same time, and fix problems before they compound. By the end of the day, the system is already tighter, not just bigger.

The unexpected shift: thinking ahead while building

This was the biggest change. Because Claude handled execution so quickly, I wasn't stuck in implementation. While something was being built, I was already thinking: is the messaging right? What's missing from this flow? What would make this feel more complete?

It creates a loop: think → build → think → build → think.

Normally, implementation breaks that loop. Here, it doesn't. And that changes how you work entirely.

Wednesday — The System Starts Compounding

By Wednesday, things sped up again, but not because the agent got faster. The system did.

  • Connectors page (fully built out)
  • Contact page
  • Product page (with different theme variations)

At this point, layouts already existed, styles were consistent, and components were reusable. Building a new page wasn't really building. It was assembling pieces that already worked.

That's when it clicked: the real leverage isn't just AI. It's pairing AI with a solid system early.

Thursday & Friday — Slowing Down Where It Actually Matters

By Thursday, most of the site was there. Not perfect, but complete enough that I wasn't thinking about structure anymore. I could finally focus on the parts that actually matter: the homepage and the product page.

These were the two pages I wanted to spend the most time on. Not just building them, but exploring them:

  • Different messaging directions
  • Different layouts
  • Different ways of structuring the story
  • Experimenting with visuals from the app to support the content

This is where the process changed again. Earlier in the week, speed was everything. Now it was about trying ideas quickly, seeing what works, and refining taste.

Because the system was already in place, I could spin up variations fast, test different approaches without friction, and throw things away without hesitation.

And that's the real advantage.

Normally, by the time you get to the homepage, you're tired. You've already spent days on setup and structure, so you just try to get it "done". This time was different. I had the space to actually think: is this message clear? Does this flow make sense? Do these visuals support the story, or distract from it?

It stopped being about building pages and became about shaping the product's story.

What Actually Changed

Looking back, the biggest shift wasn't speed. It was how I approached the work.

I stopped waiting to be "ready". I started quickly, figured things out while moving, and adjusted constantly.

I fixed things immediately. No backlog. No cleanup phase. If something felt off, I fixed it. That compounds into a much cleaner system.

I stayed in momentum the entire week. I never really got stuck. I wasn't blocked by implementation, and I wasn't overthinking decisions. It was just: keep moving.

The Tradeoff

This way of working is messy. You jump between tasks, solve things out of order, and sometimes redo work.

But most "clean process" only exists because execution is slow. When execution becomes fast, strict order matters less.

And the upside is obvious: you ship an insane amount in a short time, without losing momentum.

The Bigger Realisation

AI didn't just make me faster. It changed my role.

From someone who builds everything → to someone who decides what gets built.

That shift means the bottleneck isn't time anymore. It's taste, clarity, and decision-making.

What I'd Do Differently Next Time

  • Lock in the system even earlier
  • Be less strict about order
  • Lean harder into fixing things as I see them
  • Push further on experimentation

Because the real limit isn't how fast the agent can build. It's how fast you can decide what to build next.

Final Thought

The most surprising part wasn't the output. It was this: I never felt stuck. And more importantly, I had time to focus on the parts that actually matter.

No setup friction. No styling rabbit holes. No decision paralysis.

Just: thinking → building → improving → repeating.

And honestly, that's probably how building should feel.