The sequence diagram below traces the whole video-production chain. The
scene-inventory-extractor-v2 skill prepares source analysis,
reference images, the prompt keyword library, continuity inventory,
recurring visual element references, shot-frame assets, and its Phase 13
handoff. It does not assemble final video prompts or call Higgsfield.
shot-specifier then loads that scene pack, generates storyboard
frames through nanobanana, runs and actions storyboard consistency
checks, consults the model-specific deep-dive skills during per-shot
routing, and writes prompt files plus prompts/manifest.md.
video-generator owns the operational run from manifest to local clips:
it inspects the Higgsfield Model Context Protocol (MCP) schema, resolves
and uploads media, applies model overrides and explicit audio
preferences, submits jobs, polls status, downloads clips, updates the
generation log, and writes assembly order.
Figure 1 — End-to-end production sequence. The extractor stops after
Phase 13 and hands a checked scene pack to shot-specifier.
shot-specifier actions consistency findings before it writes prompt
files and the manifest. video-generator is the only skill that calls
Higgsfield video generation, and it must verify required references,
audio preferences, model overrides, and resumable job logging before a
shot is considered complete.