/video-generator

name: video-generator
description: >
  Execute production video generation from shot-specifier outputs through the Higgsfield
  Model Context Protocol (MCP). Use when prompts, storyboard frames, media roles, model
  routing, Higgsfield MCP uploads, generate_video calls, status polling, retakes, resume
  behaviour, or final assembly order are needed. Bridges
  structured [TAG] prompt files to model-native plain text prompts, validates model
  duration/aspect constraints, decomposes key-frame shots into supported start/end image
  clips, tracks uploaded media and job IDs, and writes generation logs.

Video Generator

Turns a completed shot-specifier handoff into generated video clips through the **Higgsfield Model Context Protocol (MCP)**. This skill begins only after `shot-specifier` Phase 8 has produced prompt files, storyboard frames, model routing, generation strategies, explicit audio-generation preferences, and `prompts/manifest.md`. `scene-inventory-extractor-v2` stops at its Phase 13 handoff and is not a direct input to production video generation.

Do not use this skill for image generation; use `nanobanana` through `scene-inventory-extractor-v2` or `shot-specifier` for that. Do not backfill missing scene analysis or shot direction here; hand back to `scene-inventory-extractor-v2` or `shot-specifier` when upstream fields are absent.

This skill owns the operational gap between "prompt written" and "clip exists on disk": resolve local paths, upload or register media, cache returned handles, submit jobs, poll until terminal status, download outputs, record retakes, and write assembly order. A shot is not complete until a local video file exists or a blocker is recorded.

The job is to be boringly competent: choose an approved Higgsfield video model, pass the right images in the right roles, submit resumable jobs, and leave enough state that a later agent can continue without guessing.

Read Before Running

• For execution validation, live-schema constraints, default overrides, and empirical

output checks, read [references/model-routing.md](./references/model-routing.md). Model choice and model-native prompt ordering are owned upstream by `shot-specifier/references/model-routing.md`; do not re-route shots here unless the manifest is being explicitly repaired.

• For Seedance 2.0-specific multimodal reference planning, prompt structure, duration

defaults, quality/speed decisions, and troubleshooting, load `seedance-2-deep-dive` whenever a job uses `seedance_2_0`.

• For Kling 3.0-specific shot structure, camera language, Elements, Motion Control,

native audio/dialogue, product prompting, and troubleshooting, load `kling-3-0-deep-dive` whenever a job uses `kling3_0`.

• For converting structured `[TAG]` prompt files into model-native prompt strings, read

[references/prompt-flattening.md](./references/prompt-flattening.md).

• For shots with key frames, read

[references/key-frame-decomposition.md](./references/key-frame-decomposition.md).

• For media uploads and job state, read

[references/media-upload-and-state.md](./references/media-upload-and-state.md).

• For what Firecrawl confirmed about the public Higgsfield MCP and SDK surfaces, read

[references/higgsfield-mcp-research.md](./references/higgsfield-mcp-research.md).

Required Inputs

• `prompts/manifest.md` from `shot-specifier` with

shot order, duration, model routing, prompt file, frame paths, aspect ratio, target resolution, resolution parameter, generation strategy, review-gate metadata, model-overrides, count, required-reference audit, and explicit audio-generation preferences.

• One prompt file per shot containing both the structured `## Prompt` block and a

`## Generation Prompt` block.

• Start and end frames for every generation clip.
• Key frames only when accompanied by a split plan; Higgsfield video generation accepts

start and end anchors, not mid-clip key-frame anchors.

• Continuity inventory or the shot-specifier reference audit when available. The

consistency pass is actionable input: any continuity-critical reference or prompt constraint it names must be present in the job or explicitly blocked.

Tool And Model Rules

Use the connected **Higgsfield MCP** for production video generation. The MCP endpoint documented by Higgsfield is `https://mcp.higgsfield.ai`; authenticate through the connector's Higgsfield account flow, then inspect the live MCP tools before submitting work.

Route to these recommended Higgsfield video models unless the live MCP schema says they are unavailable:

| Need | Preferred model | Use it for | |------|-----------------|------------| | Character identity, multiple references, hero props, recurring visual elements | Seedance 2.0 | Dialogue-adjacent shots, character-centric action, prop details, continuity-sensitive scenes | | Smooth camera motion with fewer identity constraints | Kling 3.0 | Drone POV, landscape movement, machine-vision travel, forward pushes | | Cinematic image-to-video when DoP/camera treatment is the main point | Higgsfield DoP or Cinema route when exposed | Polished camera moves from strong start images | | Provider-specific style or production requirement | Veo route when exposed and approved in the manifest | Only when the manifest names it and the MCP schema supports the needed media roles |

Do not invent model IDs. Translate manifest names such as `seedance_2_0` or `kling3_0` to the exact IDs exposed by the live MCP tool surface, record that mapping in the generation log, and stop if no matching model is available.

When the resolved model is Seedance 2.0, apply `seedance-2-deep-dive` before submission: validate the reference-file plan against live MCP limits, preserve each required reference's purpose, keep production clips within the planned duration envelope, and use draft/final quality settings deliberately instead of treating quality as a repair button.

When the resolved model is Kling 3.0, apply `kling-3-0-