The author recommends splitting game development across multiple models: Gemini-4-neon develops the plan, Claude-fable-5.2 reviews it, and GPT-6.2 Sol executes it. Any model can review the commits, and models can also review one another. Opus55 handles the frontend UI.
Suitable tasks: Implementing a game according to a plan and having other models review the commits.
Unsuitable tasks: Evaluations that require verified test results, a code repository, or a runnable game as evidence. The original post provides none of these.
Applicable model versions: GPT-6.2 Sol. The other models are recorded in the original post as Gemini-4-neon, Claude-fable-5.2, and Opus55. Their official version information is not inferred.
Applicable client, agent, or API: Not specified in the original post.
Recommended reasoning tier and parameters: The author states that GPT-6.2 Sol uses “fast ultra.” This is the routing/tier description in the original post. No API parameters or configuration details are provided.
This is not a copyable prompt. What can be reused is the division of responsibilities explicitly described in the original post:
Gemini-4-neon: Discuss the approach and develop a plan.
Claude-fable-5.2: Review the plan.
GPT-6.2 Sol (fast ultra): Execute the plan.
Any model: Review the commits. Models can also review one another.
Opus55: Build the frontend UI.
Workflow stated in the original post: First discuss the approach and develop a plan, then review the plan. GPT-6.2 Sol executes it, after which one or more models review the commits. Frontend UI is assigned to Opus55.
The original post provides no game name, prompt, code, test cases, or test results, so it does not support a claim that any game was completed or passed testing.
Editorial note (not evidence from the original post): When applying this division of responsibilities to an original game, you can run the project’s existing tests after execution and manually verify core gameplay, UI interactions, and build artifacts. Record the test commands, results, and review findings together, so model review is not mistaken for passing tests.
The original post publicly describes this as the author’s “best practice for making games with large language models right now” and lists the model roles above. This is the author’s experience-based judgment. The post includes no game example, commit history, or test evidence. “fast ultra” is also recorded only as the author’s description of their usage and does not represent independently verified API parameters.
This is a brief personal workflow recommendation, not a controlled experiment or reproducible case. It does not specify how models are called, how context is passed, how tasks are divided in detail, or what quality metrics are used. Nor does it show that this division of responsibilities outperforms a single-model workflow.
The workflow structure can be reproduced by assigning model roles as described in the original post. Users must supply the specific prompts, game task, client, and testing criteria. To evaluate GPT-6.2 Sol’s execution, retain the original plan, commit diff, and actual test results, and record them separately from the author’s experience-based description.
GPT-6.2 Sol