Skip to content

Your First "Thing": Build a World

Forget Hello World. A print statement doesn't tell you anything about what an agent is. Instead, your first task is to delegate something with real moving parts and watch the agent handle all of them.

The exercise

Start Claude Code in an empty directory:

mkdir my-world && cd my-world
claude

Then type (or invent your own variant):

> Build me a small world: a single self-contained HTML file with an
> interactive 3D scene I can explore — terrain, sky, some objects, simple
> controls. No build step, no dependencies I have to install. When it's
> done, open it in my browser.

Then watch. In the next two minutes the agent will:

  1. Plan — decide on three.js from a CDN, a single index.html, orbit or WASD controls;
  2. Write — a few hundred lines of HTML/JS in one shot;
  3. Act — open the file in your browser, by itself, with a shell command.

You now have a world. Fly around it.

The actual lesson

Now make it yours — this is the important part. Iterate conversationally:

> Make the sun set over 60 seconds, and add fog as it gets dark.

> It stutters on my laptop. Find out why and fix it.

> Add a moon with realistic orbital inclination. Label the physics
> assumptions you're making as comments.

Three things to notice while you play:

  • You never said how. You stated outcomes; the agent chose implementations. That division of labor is the core skill of agentic work — and it's a skill, because choosing the right altitude for your instructions takes practice.
  • The agent verifies. When you said "it stutters," it didn't guess — it read its own code looking for the expensive thing. Later we'll make verification rigorous with feedback loops.
  • The conversation is state. "Make it yours," "fix it" — the agent tracks referents across the session. That state is precious and finite, which is why context management gets its own section.

Why 'Build a World' and not a physics problem?

Because you already know how to judge a physics result. Here, the only judgment you can apply is does the world feel right — which forces you to practice specifying, delegating, and iterating, with zero domain knowledge to hide behind. The physics comes at 10:40.

When the seminar resumes, we'll take exactly this workflow and scale it up to research: Agentic Research Patterns →