Video skill architecture

Updated May 06, 2026

The class diagram below shows the video-production portion of the toolkit and the relationships between the skills and supporting artefacts. scene-inventory-extractor-v2 produces the scene pack, continuity inventory, prompt keyword library, recurring visual element definitions, and reference images. shot-specifier consumes that package, uses nanobanana for storyboard frames, consults the Seedance or Kling deep-dive skill when a shot is routed to that model, and emits prompt files plus the manifest. video-generator consumes that handoff and submits jobs through the Higgsfield MCP. The phoneticize skill is part of the repository but is independent of this visual generation pipeline.

All nanobanana image calls in this pipeline must request model: gemini-3-pro-image-preview. If that model is unavailable or cannot accept the reference images or character-consistency images required by the current operation, the image-generation workflow stops instead of selecting a fallback model.

uses_for_reference_images

produces

defines

uses_for_storyboards

uses_for_style_language

reads_for_constraints

consults_when_model_seedance_2_0

consults_when_model_kling_3_0

upstream_of

submits_jobs_to

uses_for_seedance_jobs

uses_for_kling_jobs

contains

produced_in_phase_2_4

Scene_inventory_extractor_v2

+run_phases()

+generate_reference_images()

+verify_consistency_phase_13()

+handoff_to_shot_specifier()

Shot_specifier

+load_scene_inventory()

+plan_shots()

+generate_storyboard_frames()

+run_storyboard_consistency_checks()

+route_models_per_shot()

+assemble_video_prompts()

+emit_prompt_manifest()

Nanobanana

+generate_image(model)

+edit_image(model)

+character_consistency(model)

+multi_image_fusion(model)

Video_generator

+load_manifest_and_prompts()

+inspect_higgsfield_mcp_schema()

+upload_media_and_cache_handles()

+submit_generate_video_jobs()

+poll_and_download_clips()

+write_generation_log()

+write_assembly_order()

Seedance_2_deep_dive

+advise_when_to_use_seedance()

+plan_multimodal_references()

+shape_seedance_prompts()

+suggest_duration_and_aspect()

+troubleshoot_seedance_failures()

Kling_3_0_deep_dive

+advise_when_to_use_kling()

+plan_shot_structure()

+plan_camera_and_motion()

+plan_elements_and_motion_control()

+shape_kling_prompts()

+troubleshoot_kling_failures()

Higgsfield_MCP

+generate_video()

+upload_media()

+get_generation_status()

Prompt_keyword_library

+provide_style_phrases()

+ensure_style_consistency()

Continuity_inventory

+list_props_and_recurring_elements()

+list_constraints()

Recurring_visual_element_ref

+element_name

+location_set

+appearance_lock

+reference_file

Figure 2 — Video skill and artefact relationships. Solid arrows show runtime dependency or handoff direction: the extractor creates continuity and recurring-element constraints, shot-specifier turns them into per-shot prompt and reference requirements, and video-generator submits only after validating those requirements against the live Higgsfield MCP. The dashed arrow records provenance: the prompt keyword library is produced by extractor Phase 2.4 and then reused downstream for style consistency.