RootGPT
I’ve really enjoyed playing the online version of the board game Root with friends over the past few months. Each faction has distinct abilities and constraints, making interactions asymmetrical. The learning curve is steep, but it’s worth it.
While learning the game, I used ChatGPT multiple times to ask about rules and strategy. Early on, it seemed helpful, but as I gained experience, its hallucinations became more apparent. The reasoning models outperform 4o but still have significant gaps.
So I tried creating a custom GPT. I uploaded the official Root rules and FAQ but realized I lacked a good way to train it on all the game’s cards. I attempted to upload images from this database, but the model hallucinated wildly, failing to distinguish between suits, crafting costs, and immediate vs. persistent effects. So I broke the problem down:
I asked o1 to generate a structured list detailing each card’s requirements and effects. Since I could only upload four images at a time (the current limit), I had to oversee and correct the output extensively. Ideally, I would have uploaded a single PDF of all the cards, but reasoning models still don’t support PDFs.
I hit OpenAI’s rate limit and tested Gemini Pro 2.0 Experimental (who names these things?) and even Deep Research. However, Gemini could only handle one image upload at a time, and Deep Research provided a report rather than the desired output—full of issues, no less. I also tried Claude 3.5 Sonnet but didn’t get far. And finally, out of curiosity, I tried DeepSeek R1. Its quality was, by far, the worst of the four. I’m genuinely shocked by the positive press it’s received lately given its actual utility. Surely it will improve, but it wasn’t even close.
Eventually, I combined results from o1 and Gemini 2.0 Pro Experimental into a decent output. I then uploaded the processed list as a .txt file into the custom GPT.
I repeated this process twice: once for the standard deck and once for the Exiles and Partisans deck.
The final product performs better than off-the-shelf ChatGPT 4o or o1, but it still hallucinates, even with explicit instructions:
IMPORTANT: Never make up rules or hallucinate. If you are unsure about an answer, tell the user instead of making something up.
Ideally, custom GPTs would support reasoning models, which could better handle the provided context.
Anyway, it was a fun side project, and I hope some people find it useful. You can check it out here.