Admittedly, the mulligan phase system prompt is the weakest part of the project. I had to add heuristics to stop the LLMs from mulliganing down to just a few ca

Admittedly, the mulligan phase system prompt is the weakest part of the project. I had to add heuristics to stop the LLMs from mulliganing down to just a few cards looking for a perfect hand. The scoring for the benchmark is mostly based on if the LLM could complete legal turns, not good turns. https://github.com/CallumFerguson/mtg-auto-deck/blob/a877c08def403be0227aa489ec4747aa5f5d1b39/mtg-auto-deck-server/llm/prompt-constants.ts#L31 https://github.com/CallumFerguson/mtg-auto-deck/blob/a877c08...