Back to Journal

AI Game Dev Adventures

Building games with artificial intelligence

Day 32025-08-18
train-game

Train Switching Game Clone - Learning from Existing Games

Attempting to clone a train switching game from Luma City app by providing screenshots and detailed requirements, discovering the challenges of game-specific AI logic.

Screenshots from Day 3
Screenshot 1 from Train Switching Game Clone - Learning from Existing GamesScreenshot 2 from Train Switching Game Clone - Learning from Existing Games

Train Switching Game Clone - Learning from Existing Games

Challenge:

My maze game requirements weren't specific enough. What if I clone an existing game I can show the AI?

Attempt 1: Train Switching Game Clone

There's a fun train game in the Luma City app - basic objective is switching train tracks to route trains to correct destinations.

💬 PROMPT
Gave it screenshot, exact game name, asked it to write detailed PRD, then said "make this game"

✅ OUTCOME

  • Visually closer to what I wanted
  • Drew basic shapes, trains moving on screen
  • BUT: Pathfinding was horribly wrong
  • Trains would try to drive straight from entry to station without following tracks
  • Spent many cycles trying to fix "you have to follow the track"
  • Would get stuck at switches, wouldn't respect switch positions

Lessons learned:

  • Again, wasn't specific enough with mechanics
  • AI seems trained on lots of development but not much game development
  • Especially Claude Sonnet struggles with game-specific logic
  • Need much more detailed instructions about specific game mechanics

Next:

Try the one-shot approach again but with even more detailed requirements

Ingredients

  • AI Assistant: Claude Sonnet - AI struggles with game-specific logic
  • Reference Game: Lumosity train switching game (screenshot provided)
  • Approach: Screenshot → PRD → "make this game" one-shot method
  • Tools: Browser-based development - Free
  • Key Challenge: Pathfinding and switch logic complexity
  • Time Investment: Multiple cycles debugging track-following behavior
  • Cost: ~$3-5 in AI credits for iterative debugging
  • Lesson: AI needs much more detailed mechanics instructions

Community & Discussion

Join the conversation through comments or share your thoughts across the web. Discuss this journal entry and share your experiences.

Comments & Discussion

Join the conversation! Comments are powered by GitHub Discussions. Share your thoughts about this journal entry.

💡 Tip: Comments are stored as GitHub Discussions, while web mentions capture shares and mentions from Twitter, Mastodon, blogs, and other platforms. Together they provide a complete picture of community engagement.