# Scene Inventory Extractor

> >

- **Type:** Skill
- **Install:** `agentstack add skill-leynos-visual-storytelling-skills-scene-inventory-extractor-v2`
- **Verified:** Yes — security-reviewed for prompt injection and unsafe behavior
- **Seller:** [leynos](https://agentstack.voostack.com/s/leynos)
- **Installs:** 0
- **Category:** [Content & Media](https://agentstack.voostack.com/c/content-and-media)
- **Latest version:** 0.1.0
- **License:** ISC
- **Upstream author:** [leynos](https://github.com/leynos)
- **Source:** https://github.com/leynos/visual-storytelling-skills/tree/main/skills/scene-inventory-extractor-v2

## Install

```sh
agentstack add skill-leynos-visual-storytelling-skills-scene-inventory-extractor-v2
```

Requires the [AgentStack CLI](https://agentstack.voostack.com/docs/cli). Works with Claude Code, Cursor, and any MCP-compatible agent.

## About

# Scene Inventory Extractor

Systematic workflow for extracting comprehensive scene inventories from narrative source
material and preparing every visual asset required for AI video generation. Continuity is
a first-class deliverable: extract story-state, dressing-state, and object-state before
writing prompts so a crew can reset scenes accurately when production shoots out of
order.

## Execution Context

This skill is designed for a command-line agent with:

- **nanobanana image generation MCP** (`generate_image`, `edit_image`,
  `character_consistency`, and `multi_image_fusion` as needed), with every image call
  explicitly using `model: gemini-3-pro-image-preview`
- **Vision capabilities** (ability to inspect generated images for consistency)
- **File system access** (structured output directories)

All image generation and verification runs **silently** — no user confirmation gates
during generation phases. Halt only on consistency failures that require human judgement.
If `gemini-3-pro-image-preview` is unavailable through the nanobanana MCP, or if it
cannot accept the reference images or character-consistency images required by the
current phase, **STOP** and report the blocker. Do not switch to another image model.

## Workflow Overview

| Phase | Name | Output |
|-------|------|--------|
| 1 | Source Analysis | Annotated reading notes |
| 2 | Creative Pillars | Visual aesthetic + storytelling style + **cinematography specification** + **prompt keyword library** |
| 3 | Narrative Spine | Structure, themes, turnpoints |
| 4 | Character Bible | Character entries with **reference-image specifications** |
| 5 | Locations Bible | Location entries with **multi-angle, multi-condition scouting specs** |
| 6 | Props Bible | Props with physical descriptions and ref-image specs |
| 7 | Scene Inventory | Per-scene breakdowns |
| 8 | Continuity Inventory | Character/location/prop state tracking across scenes |
| 9 | Shot Lists | Shot tables with full cinematography fields and duration budgets |
| 10 | Thematic Image Plan | Key narrative-beat images |
| 11 | Reference Image Generation | All character, location, and prop reference images with **video role manifest** |
| 12 | Shot-Frame Generation | Start frame, end frame, and key frames per shot |
| 13 | Consistency Verification + Handoff | Vision-based QA pass; final scene pack, frame assets, role manifest, and handoff notes for `shot-specifier` |

> **Skill chain:** `scene-inventory-extractor-v2` prepares the scene pack and image
> assets. Hand off to `shot-specifier` when per-shot direction, storyboard generation,
> and prompt manifests are needed. Hand off to `video-generator` only after prompt files,
> frame paths, model routing, and generation strategy fields exist and the goal is to
> submit jobs through the Higgsfield MCP.
>
> **Read order for reference files:** Before starting Phase 2, read
> `references/cinematography-specification.md`. Before Phase 8, read
> `references/continuity-inventory.md`. Before Phase 11, read
> `references/reference-image-guide.md`. The consistency verification procedure in
> Phase 13 is defined in `references/consistency-verification.md`. The prompt keyword
> library format is defined in `references/prompt-keyword-library.md`.
> **Downstream skill:** This skill stops at Phase 13. When full per-shot storyboarding,
> detailed actor/camera/lighting direction, model routing, prompt flattening, audio
> generation preferences, and asset-pipeline management are required, hand off to the
> `shot-specifier` skill. That skill takes the scene inventory and reference images
> produced here as its input. When video generation itself is required, use
> `video-generator`; this skill does not call Higgsfield directly.

---

## Phase 1: Source Analysis

Read the source material twice. First pass: absorb narrative arc, tone, world. Second
pass: extract with pen in hand.

### First-Pass Questions

- What is the dominant visual mood? Name it.
- Where does the camera naturally want to be? (Observational? Intimate? Surveillance?)
- What objects carry narrative weight?
- What spaces define the world?
- Who moves through this world and how do they move?
- What is the **light** doing? (Time of day, weather, artificial sources)
- What is the **sound** doing? (Ambient beds, punctuations, silence)

---

## Phase 2: Creative Pillars

> **Prerequisite:** Read `references/cinematography-specification.md` before this phase.

Extract and name the visual, storytelling, and cinematographic approaches.

### 2.1 Visual Aesthetic

```text
**Name:** {Evocative 2–4 word title}
**Definition:** {One sentence capturing look and feel}
```

Specify: Palette (3–5 named colours), Lighting rules, Texture rules, Camera grammar,
Warmth rules. See `templates/scene-inventory-template.md` for field structure.

### 2.2 Storytelling Style

```text
**Name:** {Evocative 2–4 word title}
**Definition:** {One sentence capturing narrative approach}
```

Specify: Narrative rules, Rhythm rules, Scale rules.

### 2.3 Cinematography Specification

This is **new territory** relative to pure script breakdown. Define the physical and
post-production characteristics of the image pipeline:

| Element | What to Define |
|---------|----------------|
| **Format / filmstock** | Gauge (Super 8, 16 mm, 35 mm, 65 mm, digital sensor equivalent), aspect ratio, native resolution |
| **Target resolution** | Explicit pixel dimensions for generation and delivery, e.g. `1920x1080` or `1344x768`; do not put preset labels here |
| **resolution-parameter** | Provider-facing preset label when needed, e.g. `720p` or `1080p` |
| **Grain structure** | Fine/medium/coarse; organic vs digital noise; grain response to exposure |
| **Colour process** | Photochemical reference (Kodachrome, Ektachrome, Vision3 500T) or digital LUT family |
| **Colour timing** | Overall bias (cool, warm, cross-processed); scene-specific timing rules |
| **Grading rules** | Lift/gamma/gain tendencies; crush blacks or preserve shadow detail; highlight rolloff |
| **Lens language** | Primes vs zooms; focal length range; anamorphic vs spherical; characteristic aberrations (halation, flare, bokeh shape) |
| **Depth of field rules** | When shallow, when deep; rack-focus conventions |
| **Shutter behaviour** | Shutter angle/speed rules (180° default, or intentional motion blur/strobing) |

Full specification format and examples are in `references/cinematography-specification.md`.
Use this resolution contract consistently across extractor, `shot-specifier`, and
`video-generator`: **Target resolution** is always explicit pixel dimensions, while
**resolution-parameter** carries any provider preset label.

### 2.4 Prompt Keyword Library

> **Prerequisite:** Read `references/prompt-keyword-library.md` before this sub-phase.

Immediately after completing the cinematography specification, derive a **project-level
prompt keyword library**: a canonical vocabulary of adjective phrases and art-direction
terms that reliably translate each style parameter into language video and image
generation models respond to consistently.

This library is **infrastructure**, not decoration. Every model-routed video prompt later
written by `shot-specifier` must draw from it rather than re-inventing style language
shot by shot. Inconsistent style vocabulary across shots produces visible tonal drift
even when the underlying spec is correct.

Write the library as a section in the scene inventory document and also as a standalone
file at `{project_name}_prompt_keywords.md`.

```text
### Prompt Keyword Library

* **Global style phrase:** {1–2 sentences encoding filmstock, grain, colour process}
  * Example: "35 mm print, Kodak Vision3 250D characteristics, organic fine-to-medium grain,
    accurate daylight balance, slightly desaturated shadows, gentle highlight rolloff"

* **Per-location-type vocabulary:**
  * {Location type}: {canonical phrase set for colour, light, texture, atmosphere}
    * Example exterior: "flat pewter Atlantic overcast, wet peat surfaces, no hard shadows,
      lifted blacks, fence posts leaning in wind"
    * Example interior warm: "warm amber grow-light, artificial June, high humidity haze,
      strawberry scarlet at full vibrancy"

* **Per-lighting-condition vocabulary:**
  * {Condition name}: {phrase set}
    * Example pre-dawn: "sodium-orange practicals, deep grey-blue surround, rain halos
      around light sources, near-dark ground plane"

* **Global negative constraints:** {what must never appear in prompts}
  * Example: "no trees, no dual carriageways, no US road markings, no hard shadows,
    no blue sky, no bright saturated colours except strawberry scarlet"

* **Selective saturation rules:** {objects exempt from global desaturation}
  * Example: "strawberry scarlet is exempt from the global 15–20% desaturation pass;
    push selectively against every other desaturated element"

* **POV-specific overrides:** {e.g. machine-vision shots that require no grain}
  * Example drone POV: "digital-flat, no grain, deep focus throughout, gimbal-stabilised,
    machine vision — no organic camera movement"
```

---

## Phase 3: Narrative Spine

Document timeframe, structural approach, operational themes (4–6), and key turnpoints
(3–7). Themes must be operational: "Signature authority: who gets to update 'reality'"
not "The loneliness of modern existence."

---

## Phase 4: Character Bible

For each character, extract using the appropriate template from
`templates/character-template.md` (primary / functional / interface / voice-only /
collective).

**Addition — Reference Image Specification per character:**

For every character who appears on screen, append:

```text
* **Reference image requirements:**
  * Primary: {Full body, neutral pose, front-facing, white background}
  * Face detail: {Head-and-shoulders, ¾ angle}
  * Expression set: {List key expressions needed by the story}
  * Wardrobe variants: {Each distinct outfit as a separate ref}
  * Action poses: {Any signature physical action}
```

Match detail to narrative weight: primary characters get full sets; functional characters
get primary + face; collectives get a representative figure.

Also record a continuity state chain per recurring on-screen character: wardrobe by
scene, carried items by scene, body-state continuity, and pocket / hand continuity
risks. "Same outfit" is not enough when carried objects or physical condition change.

---

## Phase 5: Locations Bible

> **Critical:** Locations receive the most comprehensive reference-image coverage. Each
> location that appears in more than one scene, or whose atmosphere shifts across the
> story, requires **multi-angle, multi-condition scouting**.

For each location, extract using `templates/location-template.md`.

### Scouting Specification

For each location, define:

| Scouting Dimension | What to Specify |
|--------------------|-----------------|
| **Angles** | Establishing wide, interior/working angle, character-entry POV, signature detail insert |
| **Lighting conditions** | Each distinct time-of-day or artificial-light state the story requires |
| **Weather conditions** | Each weather state the story requires (if exterior or weather-visible) |
| **Seasonal state** | If the story spans seasons and this location appears across them |
| **Narrative state** | If the location degrades, transforms, or is altered by story events |

The scouting matrix is: `angles × (lighting conditions × weather conditions × seasonal/narrative states)`. Generate the full cross-product only where the story **actually visits** that combination. Do not generate speculative combinations.

For recurring locations, add a continuity chain across appearances: first-appearance
state, subsequent appearance changes, and non-negotiable anchors that must never drift.

---

## Phase 6: Props Bible

For each significant prop (passes the significance test: narrative weight, multi-scene,
character-defining, symbolic, or physically interacted with on camera), extract physical
description, narrative function, sequence appearances.

For continuity, also record custody, state progression, set-down / pickup moments, and
scene exit status.

### 6.1 Recurring Visual Elements

Also identify **Recurring Visual Elements**: any object, fixture, interface, machinery,
set dressing element, furniture layout, or repeated environment component that appears
in more than two shots and has a specific enough appearance that the audience would
notice if it changed. These elements do not need full Props Bible custody/state entries
unless they are handled or narratively transformed, but they do need locked reference
images.

Treat recurring visual elements as reference-required even when they feel like location
dressing. Examples include monitor-bank layouts and screen colours, inspection robots,
grow-light strip configurations, cargo pod designs, storage cabinet designs, recurring
signage clusters, and a named character's workstation or console layout.

For each recurring visual element, record:

```text
* **Recurring visual element:** {Name}
  * Location / set: {Where it belongs}
  * Appearance lock: {Shape, layout, colour, screen state, arrangement, scale}
  * Appears in shots/scenes: {S01_SH001, S01_SH003, ...}
  * Reference image requirement: locked in Phase 11
  * Reference file: refs/recurring-elements/{name}/primary.png
  * Must pass as referenceImagePaths in: {All shots where visible}
```

**Addition — Reference Image Specification per prop:**

```text
* **Reference image requirements:**
  * Reference priority: required-before-Phase-12 / incidental
    (required-before-Phase-12: prop appears prominently on screen across multiple shots
     and its visual identity is story-critical — a named vehicle, weapon, device, or
     object whose appearance the audience will track. Generate before any location or
     scene image that includes it. incidental: prop appears briefly or in background;
     standard Phase 11 timing applies.)
  * Primary: {Object on white background, ¾ angle}
  * Detail: {Any specific detail the camera will see in ECU/INS}
  * In-context: {Object in its typical environment, if context matters}
  * State variants: {If prop changes condition across the story}
```

---

## Phase 7: Scene Inventory

For each scene (new scene on location change, significant time passage, POV shift, or
mode change):

```text
#### SC-{XX}

* **Scene ID:** SC-{XX}
* **Location:** {Location name}
* **Time:** {Time of day; weather; season}
* **Lighting condition key:** {Maps to scouting matrix entry}
* **Characters present:** {Who}
* **Objective / tension:** {What's at stake}
* **What changes:** {State shift by scene end}
* **Key sensory notes:** {Smell/sound/temperature}
* **Transitions in/out:** In: {from}. Out: {to}.
* **Continuity dressing notes:**
  * Fixed location anchors: {architecture, furniture, installed fixtures}
  * Recurring visual elements: {monitor banks, robots, light strips, cabinets, console layouts, signage clusters}
  * Movable dressing: {objects that can shift position}
  * Character-carried items: {by character}
  * Consumables / depletion states: {food, drink, cigarettes, fuel, paper stacks}
  * Weather / dirt / damage state: {mud, blood, sweat, rain, soot, wrinkles}
  * Reset-sensitive details: {what must match across coverage and return visits}
```

---

## Phase 8: Continuity Inventory

> **Prerequisite:** Read `references/continuity-inventory.md` before this phase.

Write a separate continuity deliverable at `{project_name}_continuity_inventory.md`.
Continuity extraction is pre-generation story/state tracking, not post-generation image
QA.

Assume scenes will be shot out of order. Extract continuity so that a separate crew,
working days later, can restore the exact character, prop, and dressing state without
re-reading the source material.

---

## Phase 9: Shot Lists

Create shot tables for each sequence. Use the expanded format from
`templates/shot-list-temp

…

## Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

- **Author:** [leynos](https://github.com/leynos)
- **Source:** [leynos/visual-storytelling-skills](https://github.com/leynos/visual-storytelling-skills)
- **License:** ISC

Install and usage instructions live in the source repository linked above.

## Pricing

- **Free** — Free

## Security capabilities

Automated source analysis of v0.1.0 — what this tool can access:

- **Network access:** no
- **Filesystem access:** no
- **Shell / process execution:** no
- **Environment & secrets:** no
- **Dynamic code execution:** no

*"Yes" means the capability is present in the source — more access means more to trust, not that it is unsafe.*


## Versions

- **0.1.0** — security scan: passed — Imported from the upstream source.

## Links

- Listing page: https://agentstack.voostack.com/l/skill-leynos-visual-storytelling-skills-scene-inventory-extractor-v2
- Seller: https://agentstack.voostack.com/s/leynos
- Browse the marketplace: https://agentstack.voostack.com/browse

---
Listed on AgentStack — the marketplace for AI agent skills and MCP servers. Every listing is security-reviewed. Creators keep 70%.