This started out as a casual side project, tinkering with AI to play some of my favorite old adventure games like King's Quest and Space Quest in the browser. Now it's a self-running, self-maintaining, self-iterating side project capable of importing, playing (single player or multiplayer), or modding those games (manually or with AI), or creating completely new games from scratch. It sort of invented a couple of its own "languages" along the way? It calls one of them the "machine form" which is in the ballpark of LLVM IR and ghidra p-code, and the other one is a TypeScript subset, meant to be a more human-readable layer on top of the machine form. Honestly I'm still trying to catch up a bit on what it's done and continuing to do.

I linked to the "Making Of" page of the app because I think my process and "collaboration" with Claude might actually be even more interesting than the project itself.

As a software engineer, I've been grappling with how to use or not use AI in this new world we're living in. Here's my most optimistic attitude after working on this project:

- Coding before AI was more like building with legos, one piece at a time.

- Coding with AI is more like sculpting, where you already start with material, and you shape and shave and chip off what you don't like until you get something closer to what you want.

For this project, I provided (and sometimes continue to tweak) the main goal and vision. At the beginning I still kicked off each Claude session myself and was actively prompting each step. At some point though, enough documentation had been made that my sessions started becoming one prompt of "what's next" and then a second prompt of "ok, go". So now I have an "ops coordinator" Claude session that runs routinely and spins up additional Claude sessions to keep working on the project...without me. They do things like work through the queue of already-planned work, brainstorm new items to work on, test existing work, do research into related projects or applications, have meetings with each other, and propose new directions or breaking changes to me. If there's something that requires me, I get notified. There are also several Claude html artifacts that function as a a sort of dashboard that I can pull up to see what's been done recently, what's coming up next, and what questions or action items there are for me.

Even though that sounds very hands-off, I find that I'm actually still pretty involved, because I like to redirect and nudge what the agents are working on pretty frequently. And that process does still feel pretty creative and engaging. Hence my "sculpting" attitude.

But, there are other possible metaphors for coding with AI that also come to mind...

- IKEA

- Frankenstein

So at this point, I guess it's still an open question: is what I'm making more of a sculpted work of art, or more like Frankenstein's monster?


Comments URL: https://news.ycombinator.com/item?id=50517319

Points: 1

# Comments: 3