visual-storytelling-skills

🎬 visual-storytelling-skills

Agent skills for AI film production — from prose to picture.

Every story contains a film. These skills find it. This is a collection of Claude Code skills that walk a narrative from raw prose through continuity extraction, reference-image generation, shot direction, and model-routed video-prompt assembly, with a TTS phoneticizer for good measure.

Why visual-storytelling-skills?

AI video generation is powerful but fussy. Models want precise prompts, carefully curated references, continuity-checked assets, and shot direction that could actually be handed to a camera operator. Doing that by hand for every shot in a feature is a recipe for RSI and despair.

These skills automate the production-prep pipeline:

• Continuity first: extract character, location, and prop state before a single prompt is written, so scenes shot out of order stay coherent.
• Reference-led generation: build a complete image asset library before touching a video model.
• Model-aware routing: route shots between Seedance 2.0, Kling 3.0, and whatever arrived last Tuesday based on shot type, reference availability, and budget.
• Narration-ready output: phoneticize proper nouns, brand names, and contested acronyms so your TTS engine does not mangle "Siobhán" or "df12".

Quick start

Installation

git clone https://github.com/df12-productions/visual-storytelling-skills

Typical production run

Start with a script or prose fragment. The scene-inventory-extractor parses it, extracts characters, locations, and props, and generates your full reference-image library. Its final consistency pass must be acted on before handoff: fix BLOCK findings, resolve fixable WARN findings, and turn any remaining WARN findings into explicit shot-specifier constraints.

/scene-inventory-extractor-v2

Hand the resulting inventory to the shot-specifier, which decomposes every scene into numbered shots with full directorial direction, storyboard keyframes, and generation-ready video prompts:

/shot-specifier

When the prompt manifest and storyboard frames are ready, use the video-generator to submit clips through the Higgsfield MCP, poll jobs, download takes, resume interrupted runs, and write assembly order:

/video-generator

When the clips are selected and the assembly order is ready, package them into an OpenShot project:

/media-project

When the narration script is ready, phoneticize it before sending it to Eleven v3:

/phoneticize

Image-generation tasks throughout the pipeline are handled by the nanobanana skill, which the other skills call automatically. You can also invoke it directly for standalone image work:

/nanobanana

Overall flow

flowchart TD
    prose["Script or prose fragment"]
    extractor["scene-inventory-extractor-v2"]
    inventory["Continuity inventory\nCharacters, locations, props, recurring visual elements"]
    refs["Reference image library\nCharacters, locations, props, visual elements"]
    shotSpec["shot-specifier"]
    shotPack["Shot specs\nCamera, blocking, lighting, timing"]
    frames["Storyboard frames\nStart, end, and key frames"]
    prompts["Prompt manifest\nGeneration prompts, model routing, media roles"]
    videoGen["video-generator"]
    mediaProject["media-project"]
    higgsfield["Higgsfield MCP\nSeedance 2.0, Kling 3.0, DoP/Cinema, Veo when approved"]
    clips["Generated takes\nDownloaded clips and job log"]
    assembly["Assembly order\nSelected takes and sub-clips"]
    nano["nanobanana\nGemini 3 Pro Image Preview"]

    prose --> extractor
    extractor --> inventory
    extractor --> refs
    extractor -. "image prompts and consistency refs" .-> nano
    nano -. "locked reference assets" .-> refs

    inventory --> shotSpec
    refs --> shotSpec
    shotSpec --> shotPack
    shotSpec --> frames
    shotSpec --> prompts
    shotSpec -. "storyboard frame generation and edits" .-> nano
    nano -. "start, end, and key frames" .-> frames

    shotPack --> videoGen
    frames --> videoGen
    prompts --> videoGen
    refs --> videoGen
    videoGen --> higgsfield
    higgsfield --> clips
    clips --> assembly
    assembly --> mediaProject

Skills

Image-generation skills must use the nanobanana MCP with model: gemini-3-pro-image-preview. If that model is unavailable or cannot accept the required reference or character-consistency images, the workflow stops instead of falling back to another image model.

Learn more

• Users' guide — workflow diagrams and architecture overview
• AGENTS.md — agent workflow guidelines and conventions
• skills/*/SKILL.md — full workflow documentation for each skill
• skills/*/references/ — supporting reference documents loaded automatically at the relevant phase of each workflow
• notes/ — research logs, model-routing evidence, and open questions

Licence

ISC — see LICENSE for details.

Contributing

Contributions welcome. See AGENTS.md for conventions and workflow guidelines.