The official release page publishes interaction-prototype prompts that can be copied directly. They are suitable for using Terra to generate a runnable frontend draft first, then refining its visuals and interactions through rendering checks and iteration.
Suitable tasks: Interactive explanations, frontend visualizations, lightweight game prototypes, and ChatGPT Work/Codex workflows that require checking rendered results.
Unsuitable tasks: Treating the example prompts directly as production requirements, designing backend architectures, or accepting safety-critical code.
Applicable model versions: The GPT-5.6 series; the release page does not label these examples as Terra-specific, so Terra's actual performance requires independent retesting.
Applicable clients, agents, or APIs: The release page demonstrates the interactive experience in ChatGPT Work; on the API side, the corresponding tools must be integrated separately.
Recommended reasoning tier and parameters: Start with the default/medium reasoning setting to complete the first version; if max is available, enable it only for complex interactions and final acceptance, and record cost and latency separately.
The 3D prototype prompt published on the official page:
Can you implement a 3d sailing game for me? For anything that needs bitmaps/textures/sprites (or if it helps to have a mockup reference for any 3d models you build) feel free to use imagegen.The interactive-explanation prompt published on the official page:
Create an interactive spirograph to explain how it works.Copy one prompt first, then add the target users, screen size, interaction method, completion criteria, and dependencies that must not be introduced.
Have the model deliver a runnable minimum version first; do not request all visual details at once. Record the model, reasoning tier, tools, and generation time.
If the task requires bitmaps, textures, or sprites, allow the image-generation tool and include the purpose of the assets in the acceptance checklist.
Run the prototype in the browser or client and check loading, input responsiveness, layout, readability, and error logs; report failed behaviors exactly as observed.
In the second round, address only reproducible issues, and require the model to list changed files, verification commands, and the parts that remain unverified.
In the final round, have the model make visual corrections based on the rendered result; do not treat “code was generated” as “interaction acceptance passed.”
The release page describes GPT-5.6 as capable of creating usable interfaces from high-level guidance and checking and optimizing the rendered result.
The page displays two complete prompts, corresponding respectively to a 3D sailing game and an interactive explanation of a spirograph.
The release page explicitly mentions that GPT-5.6's computer-use capabilities can be used to check and optimize rendered results. This supports a “generate—run—check—iterate” workflow rather than guaranteeing success from a single generation.
The release page also introduces max and ultra reasoning/multi-agent settings, but does not provide a Terra-specific frontend acceptance score.
The examples come from a GPT-5.6 series release demonstration, not a blind test; the random seed, complete environment, dependency versions, and acceptance scripts have not been published.
“Runnable” is not the same acceptance criterion as production-grade security, accessibility, performance, and copyright compliance; these must be checked separately.
Whether the image-generation tool is available depends on the client and permissions; an API workflow cannot assume it is exactly the same as ChatGPT Work.
The release page does not publish Terra-specific duration, cost, or success rate for the examples, so they cannot be extrapolated into quantitative commitments for Terra.
The original prompts provided on the page are short sentences. Their key value is that they put the goal, optional asset tools, and final interactive object in the same request.
GPT-5.6 Terra