Prompt packs for generated shots

prompt_pack compiles a shot card into each model family's prompt syntax. It is a prompt package: no generation, no spend and no network call. Say that plainly to the user, and never describe a pack as a render or a clip. Paid generation arrives in Phase 7, behind the engine's policy, spend limit and consent gates.

1. Write the shot cards

Each generated_video or avatar scene gets a shot card in project/video-spec.json. The card never names a provider or a model. Direct it like a shot list:

• purpose: the shot's one job (emotion, plot or pressure).
• action: one clear action by the subject ("the engineer leans back and exhales"). A second action belongs in its own shot.
• camera: one move ("slow push in") or locked. "Push in and then pan" is two shots.
• subjects: stable ids bound to reference assets ({id, role, asset}), where the role says what the reference is for ("identity, wardrobe").
• environment (place, time, one detail), look (lens, light, grade) and audio (dialogue by subject id, at most 3 SFX tied to visible events, ambience).
• on_screen_text stays post. Logos, prices, UI and copy are composited in the edit, never generated, so leave them out of the action.
• continuity, end_state and first_frame_from (the previous scene id) chain one shot into the next.
• exclusions: what must not happen. They become a negative prompt only where the model accepts one. Runway gets them as positive statements.

Run spec_validate first.

2. Compile

Call mcp__plugin_video-studio_engine__prompt_pack {project_dir}. It compiles for all six families by default. Narrow it with families: ["veo", "kling"] or scenes: ["s03"]. It writes:

• prompts/<family>/<scene>.md: the prompt, provider-neutral parameters, warnings with fixes, what the compiler adjusted, and notes;
• prompts/<family>/<scene>.json: the same data, for a Phase 7 adapter;
• prompts/README.md: the index, the consistency plan, and for each family which credential Phase 7 would need and whether it is set. Only a yes/no is recorded, never the value.

3. Read the warnings and fix the card

Every warning has a fix. The common ones:

• camera_compound, action_compound: split the shot.
• brand_text_in_post: describe a clean surface and add the text in the edit.
• duration: the provider only takes certain lengths. The pack snaps the length; trim or hold in the edit, or split a long scene into chained shots.
• aspect_ratio: the model lacks the spec's ratio. Reframe in the edit or use another family.
• cast_limit, cast_unbound: too many references, or a model that cannot bind them. Chain from an identity keyframe instead.
• audio_not_native: add the sound in the edit.

Edit the card, not the generated .md, and run prompt_pack again. The output is deterministic: the same card always gives the same text.

4. Consistency plan (tell the user)

1. Identity keyframe first. Approve one still of each recurring subject before any clip is made. It is the reference every shot binds to.
2. Riskiest shot first. Generate the hardest shot (hands, faces, dialogue, fast motion) and inspect it before paying for the batch.
3. Chain the shots. Export each approved shot's last frame to prompts/frames/<scene>-last.png. A shot with first_frame_from starts from it, and its params already point there.
4. Logos, prices and copy in post. Composite them in the edit over clean generated plates.

5. Be honest about the specs

Every provider spec (provider-specs/*.yaml) is a dated hypothesis with verified: false, and every pack repeats that note. Tell the user to check the provider's live docs before generating. To re-verify a spec, see docs/contributing/providers.md. Sora is not supported: its API was shut down on 2026-09-24.