A developer reran a one‑shot game prompt that had originally been produced with GPT‑3 and DALL‑E four years ago, this time using Codex Desktop running GPT‑5.6 Sol Ultra — the Sol mode that makes aggressive use of sub‑agents. The result was a more elaborate playable game titled Moonlight & Mayhem.
How this version differs from the previous run
The earlier run with Claude Fable 5 produced a simple backyard game where a single raccoon collected coins and fish. GPT‑5.6 Sol Ultra’s version relocates the action to a museum: the player must rescue two raccoon crewmates, stack on top of them, and break a golden sardine out of its display case — a clearer heist scenario.
Assets and tooling
The project’s GitHub repository includes the textures and prompts generated by the model; textures were created using gpt-image-2. The developer also published the full Codex transcript in the repository.
The bug and its fix
The one‑shot output initially contained a visible bug: each raccoon had a huge black sphere — resembling an oversized eyeball — floating above it. Although the developer reviewed screenshots during development, Codex did not automatically detect or correct the issue. The bug was removed after two follow‑up prompts:
- "Why do the raccoons have huge black spheres on them?"
- "Fix it"
Those prompts produced a fix; the corrected version is also available in the repository.
Time and cost
Codex completed the session in 52 minutes. The developer included an AgentsView cost estimate showing what the session would have cost under full API pricing rather than using their monthly Codex subscription.
Why this matters
The experiment illustrates two points: larger agent‑oriented models can generate more sophisticated, playable results and associated assets, but generated outputs can still exhibit obvious visual or logical errors that require human prompting or oversight to correct. The shared assets and transcripts make the run reproducible for others who want to inspect the model’s outputs or try the game.
Tags: game‑design, ai, openai, generative‑ai, llms, coding‑agents, codex, gpt
