I started by thinking about the game mechanics I actually wanted to see. I come from a long background in diceless tabletop RPGs, and what I loved about them was the GM improvising, making rulings on the spot. When LLMs came along, I saw a chance to build a mechanic that worked like that.
But LLMs on their own are bad at it. They regress toward the mean and they try too hard to please the player. Nobody ever really fails. So I built a system (I won't name it, I'm not here to self-promote) around a 3d10 bell curve of outcomes. The dice decide how well an attempt goes, across a whole spread of grades from disastrous to legendary, and the LLM has to narrate whatever the dice say. That forces it to produce results the player won't like, which is where the drama comes from.
With that in hand, I could resolve any conflict narratively. Two characters chasing the same goal, one trying to kill or rob the other, whatever. The roll leans the story in one direction or the other. Then I realized I needed the LLM to be able to add to or subtract from the roll itself, so I built a classifier of sorts: it looks at the situation, decides who has advantages and who has disadvantages, and those become modifiers on the dice.
The last piece was making the story actually matter. On the back end, I take the narrated outcome apart and figure out which parts of the game state it should change, so the fiction and the mechanics stay in sync.
The end result plays like Game of Thrones: scheming, backstabbing, and assassinating your way to power in a kingdom. That's just the first mode, basically a conquest mode where you take over the kingdom. Since resolution is story-based, other modes are really just different stories to resolve, so more are planned.
The system behind all this goes back years. The game itself is finally done.
Comments URL: [https://news.ycombinator.com/item?id=48998591](https://news.ycombinator.com/item?id=48998591)
Points: 1