AI experiment and conclusion

Overview

The purpose of this site and related games was to gauge the capabilities of current AI tools.
Before this experiment, I was not convinced AI is a threat to general development. My opinion has changed.

Nothing about this project was original (except the name and the character).

I started out with literally the default prompt suggestion on Replit. The entire concept was not mine.

Within 5 minutes, about 20% of what is visible was ready. The rest took about 40-50 hours to refine for production. By "refine" I mean a cycle of prompts to fix UI/UX and backend issues like auth and persistence.

Surprises/what is AI good at?

Backend

To my surprise, Replit generated databases, logins, security that seems to work, game persistence and everything needed to build these games. All games except crossword was almost a one-shot prompt requiring very little guidance.

I did not expect the competent backend capability, built on node and Postgres.

I made Replit also assume an identity chosen by itself (Nova). Devlogs on the site were also written by Nova.

Simple game logic

Replit nailed game logic one shot on all games except crossword. It also designed and built the concept and UI almost always on the first attempt.

Code quality

The code was surprisingly well structured and documented in an updated readme file.

CI/CD and deployments

Everything is automated. It's a 1-click process to get to prod after the AI does its thing.

Challenges

UX 

AI sucks at this. It requires significant guidance and iteration to get this right. Button placement for example.

Text input constantly getting stuck behind mobile keyboards is another example.

Complex algorithms

Crossword was a challenge to get done without writing code. I ended up need a "collaberation" between ChatGPT 5.1 and Replit, ultimately resulting in what I still believe is a sub-par crossword experience.

Actual games

After seeing the initial capability, I tried using these tools to generate "actual" games. Something like a vertical mining type of game. Something simple where the player mines downward for ores, buys better tools, repeat.

This was almost impossible. It seems incapable of reasoning about "world space" and input nuances.

Polish

Initial output requires many many iterations to get to a production ready system that is reasonable well structured and functional. Many times the output does not work at all, despite automated test tools.

Cost

From start to finish, this cost about $150-$200 to get done using only AI.

Conclusion

For sites and systems that are not art/graphics intense, and have simple algorithms, I recommend and fear these tools. The output was surprisingly good. I've since started using these to prototype for work projects as it works 90% of the way for business type of scenarios where the logic does not have many creative components.

For production-ready 2D or 3D games, I don't think these tools are anywhere close to achieving this capability. As soon as spatial reasoning becomes valid, the tool falls apart. It's definitely more suited to websites and word games. Potentially simple spatial games like snake.

Overall AI was terrifyingly good at building the majority of Shikou, which I believe is of reasonable quality to run as a production site. It does require significant human guidance and curation of content and usability aspects.