Studio uses a language model to write. It does not use a video-diffusion model in the first milestone.
Language model#
The compiler sends one prompt and expects a structured plan: script, shots, and composition source that matches the selected skill. The default adapter is local-planner. It builds that plan from the brief so a fixture can render with no API key. A response that is not a plan fails the job and stores the raw text. Claude and OpenAI stay later adapters. The provider is a setting, not a fork of the pipeline.
The call stores the model id and the prompt. A later regeneration uses the same skill even if the default model changes.
What the first milestone does not call#
Pixel video models (local Wan, Hunyuan, LTX, CogVideo, or hosted Veo, Kling, Runway, Sora) stay off the default path. They are the right tool for a photographed look, and the wrong default for a product that must render without a GPU and without a per-second video bill.
docs/web-research.md records when to add one.
Grounding#
If the brief contains facts, the research step has to fetch a page and keep a quote. The model does not invent a benchmark, a price, or a date. The fact-check skill compares the spoken script with that quote list.