GPT-6 Astra workflow builds a car site without 3D models

• The car configurator swaps 3D assets for generated visual states: AmirMushich’s demo post says LTX API made the video transitions and GPT Images 2.5 handled car colors and wheels.
• Video generation happens during asset preparation, while the browser plays prepared clips during interaction, according to AmirMushich’s workflow post.
• The build is unusually reproducible for a flashy web demo: AmirMushich’s repository follow-up links the code, while his prompt post publishes a copy-pasteable build prompt.
• The creator split agent effort by task, using ExtraHigh for planning and edits, and High for coding in AmirMushich’s settings reply.

The LTX text-to-video API returns a job ID before a finished clip, then exposes the result after polling. A related AmirMushich walkthrough treats those clips as the website’s transition system, rather than decoration around the interface.

Mushich’s post assigns separate jobs to three parts of the stack:

• GPT-6 Astra: site design.
• A coding agent connected to LTX API: video-transition generation.
• GPT Images 2.5: color and wheel selection.

He also says the demo used no 3D models.

AmirMushich’s [linked Video States Website repository]Video States Website repository uses pre-generated forward and reverse clips, with no live video generation during user interaction. The project describes this as a seam-safe playback system, and separates asset production from the website build.

LTX’s official endpoint accepts a text prompt describing scene, camera motion, and mood; it responds with a job ID, then returns the video URL once the job is complete. That is the generation path the coding agent is wired to in Mushich’s demo.

Mushich distributed the handoff across four posts:

• The code repository is linked in AmirMushich’s repository follow-up.
• The complete build prompt appears in his prompt post.
• The LTX connection setup appears in his API setup post.
• The ZIP and prompt bundle appears in his project download post.

The linked repository is a production resource pack, not a root-level app that runs with a single npm start command. Its build brief asks a coding agent to create the application in a separate folder, use the supplied public assets, and test the state cycles through a preview URL.

Mushich said in his settings reply that he divided reasoning effort by stage:

• ExtraHigh: planning.
• High: coding.
• ExtraHigh: edits.

He also said he had been using the LTX API through Codex lately in another AmirMushich reply. OpenAI’s Astra prompting guidance similarly identifies task prompts, AGENTS.md, and skills as instruction layers that warrant revisiting with the new model.

Before generating motion, Mushich’s companion walkthrough planned a single composition that could carry every state:

• a 16:9 frame;
• a completely static camera;
• a full-body subject centered in frame;
• architecture on both sides; and
• a central 9:16 safe area for mobile.

Mushich’s walkthrough assigns each destination state a forward clip and a reverse clip that returns to the identical base frame. Switching branches begins from that base, keeping the transitions aligned rather than asking the video model to improvise continuity between every possible pair.