Word Bullet Hell
Catch word bullets and form sentences
Overview
You play as someone trying to navigate through life while struggling to put your thoughts together. Whether it is trying to find love or convince your boss of a raise, communicating can be difficult. Your brain can easily become spaghetti. In this chaotic word bullet hell, you will be trying to formulate coherent thoughts by determining which words you want to get hit by to build your thoughts and ideas.
Play on Itch.io
Gameplay
The player navigates a screen filled with word and emoji projectiles.
- Sentence Building: Instead of typing, players must physically collide with specific “word bullets” to fill their response slots.
- Fragment-Based Input: Once the slots are filled, the game sends these fragments to an LLM “translation layer”, or the “Caveman”.
- Punctuation Difference: Punctuation would be very important in how the conversions would occur. For example, “me hungry!” would turn into “I am super hungry,” versus “me hungry?” could turn into “Am I hungry?”
Design Goals & Iterative Process
The primary goal was to create a seamless bridge between non-linear, segmented gameplay actions and structured linguistic output, balancing player agency with the divergent possibilities of generative AI.
Initially, the Caveman LLM (responsible for fragment reconstruction) and the Chat LLM operated as independent agents. While the Caveman effectively translated fragments into readable sentences, the output lacked contextual relevance to the broader conversation, resulting in disjointed narrative flow.
To resolve the disconnect, I modified the architecture to provide the Caveman LLM with access to the global chat history. This allowed the agent to synthesize fragments into sentences that were not only grammatically correct but also contextually relevant to the current scenario, ensuring that player choices felt grounded in the ongoing dialogue.
Because the conversation could branch into unpredictable territory, a static word list was insufficient for the game’s growing narrative complexity. I developed a Word Generation Engine that acts as a real-time summarization agent. It analyzes the current narrative state and dynamically populates the “bullet hell” with relevant Word objects, ensuring that the projectiles remain mechanically useful and narratively pertinent as the plot evolves.
LLM Engineering
I architected a modular, three-tier pipeline to manage the transformation of player actions into game data:
Caveman:
- This would be utilized to take the fragmented sentence the player has created and then convert it into a proper sentence to be utilized by the player.
Chat:
- This would be utilized to generate the conversation between the player and the NPC. It would take in the string of the player’s sentence, alongside the chat history of the previous conversation, and then return a string of the NPC’s response.
Word Generation:
- As players talk to NPCs, plots can go into totally unexpected routes. To ensure word bullets are still relevant and can push the conversation further, we need this Word Generation Engine to provide more words.
- It will take in descriptions of the scenario the player is in, and all relevant information about it. Then, output a list of relevant Word objects that are helpful for future conversations.