How to End Prompt Lottery and Score Inflation with a Mechanic‑Driven AI Game Iteration Framework

Generating a simple game with a large language model (LLM) is now easy, but the resulting games often lack novel mechanics and suffer from "prompt lottery" and "score inflation"—the AI may produce superficial variations or give overly generous self‑evaluations.

CreativeGame, a technical report titled CreativeGame: Toward Mechanic‑Aware Creative Game Generation (arXiv:2604.19926), proposes a fundamentally different approach: treat game generation as a memory‑rich, planned, and verifiable evolution of mechanics rather than a one‑shot prompt.

The framework forces the AI to write a structured design document before any code, specifying which core loop to retain, which new rules to add, which old elements to remove, and how to reorganize existing components. This design contract is attached to four subsequent generation stages—Skeleton, Feature, Visual, and Refinement—ensuring that code implements the intended mechanics instead of merely piling up code.

For example, when tasked with creating a tower‑defense game inspired by Plants vs. Zombies , the system first extracts core elements such as Bloom (energy producer) and Shooter (75 Energy projectile). It then iteratively adds waves, curved paths that slow zombies, and a mechanic where shooters can charge sun‑cannons, turning a simple defense into a richer strategic experience.

CreativeGame also introduces CreativeProxyReward , a deterministic reward system that evaluates generated games on structural mechanism changes, planning visibility, novelty against a global expert mechanism library, and runtime robustness. Two hard thresholds reject games that fail to run or contain basic static‑analysis errors such as missing game loops or mismatched brackets.

To address the forgetting problem of single‑generation pipelines, CreativeGame implements a Lineage‑Aware Memory architecture. Each game version becomes a node in a genealogy tree, sharing a memory pool within the same lineage while isolating across lineages. When generating a new version, the agent queries both the global mechanism repository and the accumulated successes and failures of its own lineage, enabling consistent evolution of core mechanics.

Four case studies illustrate the framework’s capabilities:

Memory Relay (based on Fireboy & Watergirl ) reinterprets dual‑character cooperation as cooperation between the current player avatar, a stored past avatar, and recorded movement ghosts, allowing past actions to influence present puzzles.

Rhythm Bird (derived from Flappy Bird ) adds a beat‑synchronized ring, energy charging, a Phase state, and a "perfect pass" mechanic that reshapes future gate placements, plus a death‑echo that retroactively widens gates.

Ritual Ink Cup (inspired by Happy Glass ) replaces static ink lines with programmable ink types— Solid, Absorb, and Ritual —introducing stateful, memory‑aware water flow.

Resonance Garden (a twist on Plants vs. Zombies ) turns friendly fire into a resource: Generator blocks can store intercepted shots as charge, later released for stronger attacks, and adds lane‑bending to redirect enemy paths.

The authors argue that true AI creativity requires moving beyond one‑off, high‑variance outputs toward a structured, traceable, and iteratively improvable design process. By making mechanisms explicit, verifiable, and inheritable, CreativeGame demonstrates a pathway for LLMs to evolve from "high‑probability imitators" to genuine creative designers.