Shot Specifier

Takes a completed scene inventory and reference image set as input and produces generation-ready shot specifications: numbered shots with full directorial direction, storyboard keyframes, video prompts, model routing, and an asset pipeline.

This skill is explicitly expensive. Generating storyboard images for every shot in every scene requires many image generation calls. Per-shot video generation is more expensive still. This cost is the price of consistency and production quality — do not skip shots or reduce keyframe coverage to save cost unless the user explicitly requests it.

Input Requirements

Before beginning, verify the following inputs are available:

If the prompt keyword library or video role manifest are absent, run scene-inventory-extractor-v2 phases 2.4 and 11.6 first before proceeding.

Execution Context

This skill requires:

• nanobanana image generation Model Context Protocol (MCP) (see

references/storyboard-generation.md), with every storyboard image call explicitly using model: gemini-3-pro-image-preview

• Vision capabilities (for storyboard consistency verification)
• File system access (structured output directories)

Generation runs silently — no user confirmation gates during storyboard or prompt phases. Halt only on consistency failures that require human judgement. If gemini-3-pro-image-preview is unavailable through nanobanana, or if it cannot accept the required reference images or character-consistency images for the current storyboard operation, STOP and report the blocker. Do not continue with a fallback image model.

Workflow Overview

Read order: Before Phase 4, read references/shot-direction.md. Before Phase 5, read references/storyboard-generation.md. Before Phase 7, read references/model-routing.md. If Phase 7 routes any shot to seedance_2_0, also load seedance-2-deep-dive before finalizing the prompt, reference plan, duration, aspect ratio, and generation notes. If Phase 7 routes any shot to kling3_0, also load kling-3-0-deep-dive before finalizing shot structure, camera motion, Elements or Motion Control references, duration, aspect ratio, audio notes, and generation notes. The asset pipeline conventions are in references/asset-pipeline.md.

Phase 1: Input Audit

1. Locate and confirm the scene inventory document.
2. Inventory all available reference images and note their categories (characters,

locations, props, style anchors).

1. Load the video role manifest. If absent, halt and instruct the user to run

scene-inventory-extractor-v2 Phase 11.6 before proceeding; do not derive a replacement manifest in this skill.

1. Load the prompt keyword library. If absent, halt and instruct the user to run

scene-inventory-extractor-v2 Phase 2.4 before proceeding.

1. Load the continuity inventory if present. Flag if absent — continuity gaps will be

carried forward as WARN items.

1. Record the list of scenes to be processed and their target clip counts.

Phase 2: Shot Decomposition

For each scene in the scene inventory, decompose it into numbered shots.

Duration Budget

Before writing any shot rows, establish a duration budget:

* **Scene ID:** SC-{XX}
* **Scene description:** {brief}
* **Total duration target:** {N} seconds
* **Clip count:** {N} shots
* **Per-clip allocation:** SH001: {N}s, SH002: {N}s, ...

Budget rules:

• Clips must be 4, 6, or 8 seconds.
• One clip = one action in one location. If the action requires more than 8 seconds,

split it.

• Multi-beat scenes need multiple clips. A scene with setup + action + reaction is three

clips minimum.

• Establishing shots: 6–8 seconds. Insert/detail shots: 4 seconds. Dialogue beats: 6

seconds. Complex action: 8 seconds.

Shot Row Format

For each shot, record:

#### {S{XX}_SH{XXX}}

| Field | Value |
|-------|-------|
| Shot ID | S{XX}_SH{XXX} |
| Frame size | XW / W / M / CU / ECU / POV / INS / OTS |
| Lens | {focal length} mm {spherical/anamorphic} |
| Camera mount | {tripod / Steadicam / handheld / crane / gimbal / drone} |
| Camera motion | {static / pan / tilt / dolly / arc / zoom / rise} |
| Visual action | {Concrete, filmable description} |
| Duration | {4 / 6 / 8} seconds |
| Pacing | {slow / moderate / fast} |
| Clip boundary | {continuous / scene_cut} |
| Start frame ref | {Ref ID from video role manifest} |
| End frame ref | {Ref ID or "generate-new"} |
| Grain/grade override | {override or "global spec"} |

Phase 3: Frame Role Assignment

For each shot, assign specific reference images to video model input roles.

Use the video role manifest from the scene inventory as the primary source. Where a reference image does not yet exist for the required role, flag it as MISSING and note what needs to be generated.

### S{XX}_SH{XXX} Frame Roles

| Role | File | Notes |
|------|------|-------|
| start_image | refs/locations/launch-strip/low-pre-dawn-rain.png | Pre-launch state |
| end_image | MISSING — generate in Phase 5 | Gannet mid-lift |
| image (subject ref) | refs/props/gannet-uav/primary.png | Identity anchor |
| image (style ref) | refs/style/style_anchor_01.png | Global look |

Missing role resolution: If a start_image or end_image is missing from the reference library, generate it in Phase 5 (Storyboard Generation) as a storyboard frame, then promote that generated image to a reference role for the video generation call.

Phase 4: Shot Direction

For each shot, write full directorial direction covering:

4.1 Actor Direction

* **Actor(s):** {Who is in frame}
* **Position:** {Where in frame — left/centre/right; near/mid/far; above/below eyeline}
* **Posture/body state:** {Standing, seated, leaning; tension or ease in body}
* **Action:** {Exact physical action — what they do, in what order, at what speed}
* **Eyeline:** {Where they look — at camera, at another character, off-frame}
* **Expression:** {Named expression from character bible or described precisely}
* **Continuity state entering scene:** {Wardrobe, carried items, body state}
* **Continuity state exiting scene:** {Any changes to the above}

For shots with no human subjects (establishing shots, drone POV, inserts), mark actor direction as N/A.

4.2 Camera Direction

* **Mount:** {Tripod / Steadicam / handheld / crane / drone gimbal}
* **Starting position:** {Physical location of camera at shot start}
* **Motion path:** {Exact description of camera movement — direction, speed, arc}
* **Ending position:** {Physical location at shot end}
* **Focal length:** {mm}
* **Depth of field:** {Shallow / moderate / deep; what is sharp, what is soft}
* **Shutter:** {180° default or override with reason}

4.3 Lighting Direction

* **Key source:** {Natural / artificial; direction; quality — hard/soft/diffused}
* **Fill:** {Fill source or none; ratio to key}
* **Practicals:** {Any visible light sources in frame — colour, position, intensity}
* **Colour temperature:** {Draw from prompt keyword library for this location type}
* **Grade note:** {Scene-specific override from global cinematography spec, if any}

4.4 Effects Direction

* **Practical effects:** {Rain, smoke, steam, dust, water, fire — present/absent}
* **Environmental motion:** {Wind in grasses, ripples on water, steam from vents}
* **Subject effects:** {Rotors spinning, droplets on fuselage, breath condensing}

4.5 Audio Direction

* **Bed:** {Constant ambient — wind, hum, traffic, crowd}
* **Punctuations:** {Specific sounds at specific moments in the clip}
* **Dialogue:** {Exact words if on-screen lip-sync required; "off-screen" otherwise}
* **Music:** {Diegetic / non-diegetic / none}
* **Audio generation preferences:** ambient={on/off}; sfx={on/off}; dialogue={on/off};
  music=off; narration=off; preserve_silence={true/false}

Narration is always a separate process. Do not request generated narration, voice-over, or off-screen explainer audio in the video model prompt. Non-diegetic music defaults to off unless the user or scene explicitly requires it.

Phase 5: Storyboard Generation

Prerequisite: Read references/storyboard-generation.md.

Generate storyboard keyframe images for every shot. Use the nanobanana MCP tools with model: gemini-3-pro-image-preview on every image call.

Pre-Generation Reference Check (per shot)

Before generating any frame for a shot, verify:

1. A canonical reference image exists for every named character present in the shot.
2. A canonical reference image exists for every named prop visible in the shot.
3. A locked reference image exists for every recurring visual element visible in

the shot: fixtures, interfaces, machinery, furniture layouts, signage clusters, or set dressing that appears in more than two shots and would be noticed if it changed.

1. A canonical reference image exists for the specific location variant (angle ×

lighting condition) the shot requires.

If any is missing, generate it now before proceeding. This check is mandatory — a frame generated without a prop, recurring visual element, or location reference will invent the object or set dressing independently, producing different monitor layouts, fixture arrangements, robots, cabinets, or prop designs across the sequence.

For any shot likely to route to Kling 3.0, treat the start and end storyboard frames as the primary continuity carrier. The current Higgsfield MCP surface observed in S01 only accepted start_image and end_image for Kling jobs, with no generic image role for style, location, character, or prop references. Bake every continuity-critical character, prop, recurring element, location variant, and style constraint into the start/end frames now; do not assume video-generator can upload those refs later.

Generation Order

1. Start frames for all shots in the sequence.
2. End frames that require generation (not reused from scouting refs).
3. Key frames for shots with intermediate states specified.

Start Frame Prompt Construction

Combine:

• Global style phrase from the prompt keyword library
• Location vocabulary for this shot's location type
• Lighting condition vocabulary for this shot's lighting condition
• Actor direction (position, action, expression, outfit)
• Camera direction (framing, angle, depth of field)
• Effects direction (rain, steam, etc.)
• Global negative constraints

Reference images to pass:

• For character-centric shots, use nanobanana MCP character_consistency and set

referenceImagePaths to the character identity ref first, followed by the matching location ref, significant prop refs, visible recurring visual element refs, and style anchor when available.

• For environment or prop-led shots, use nanobanana MCP generate_image and set

referenceImagePaths to the matching location ref plus the specific prop refs needed in the frame plus any visible recurring visual element refs. Add the style anchor when available; add character refs only for visible named characters whose identity must be constrained.

Prompt ending: "no text, no watermarks, no logos, no labels, no annotations"

End Frame Prompt Construction

If the end frame shows the same subject in a significantly different state:

• Use edit_image (edit from start frame)
• Set referenceImagePaths to the start frame plus only the Phase 11 refs needed for

the described change

• Describe only what changes; do not repeat unchanged elements
• Prefer this path for any end frame derived from the start frame; character, location,

prop, and style consistency should inherit from the start frame naturally.

If the end frame shows a different composition, angle, or subject configuration:

• Use generate_image (generate new)
• Set referenceImagePaths to the start frame, matching location ref, required prop

refs, visible recurring visual element refs, and style anchor when available

• Full prompt as for start frame but with end-state description

File Naming

shots/{shot_id}/start.png
shots/{shot_id}/end.png
shots/{shot_id}/key{NN}.png

Phase 6: Storyboard Consistency Check

After generating all storyboard frames, run a vision-based consistency pass.

Checks

Output

Write a storyboard consistency report:

reports/storyboard_consistency_report.md

BLOCK-level issues must be resolved before Phase 7. WARN-level issues are logged for human review only when they cannot be fixed locally. The consistency pass is not informational: fix clear WARN issues now, and convert remaining WARN items into explicit constraints in the affected prompt and manifest row.

Phase 7: Video Prompt Assembly

Prerequisite: Read references/model-routing.md. Phase 6 complete with no unresolved BLOCK issues.

For each shot, assemble a complete video generation prompt. This skill prepares the handoff package for video-generator; it does not submit Higgsfield jobs itself.

If the recommended model is seedance_2_0, apply seedance-2-deep-dive before writing the prompt. Record why Seedance is the right route, keep the clip short unless the shot is a deliberate hero test, assign each reference file a clear purpose, and shape the prompt as a Seedance-native multimodal direction rather than a generic video prompt.

If the recommended model is kling3_0, apply kling-3-0-deep-dive before writing the prompt. Record why Kling is the right route, structure multi-beat work as labelled shots with durations, specify camera behaviour and endpoint, identify any Elements or Motion Control references, and treat critical text/UI as baked-frame or post-production work. When the live Kling route has not been proven to accept generic references, mark continuity refs in the Reference Audit as baked into start/end frames rather than as upload requirements.

Prompt Template

# {S{XX}_SH{XXX}} — Video Prompt

## Metadata
- **Shot ID:** S{XX}_SH{XXX}
- **Scene:** SC-{XX} — {scene name}
- **Duration:** {4 / 6 / 8} seconds
- **Pacing:** {slow / moderate / fast}
- **Clip boundary (next):** {continuous / scene_cut}
- **Recommended model:** {model ID — see references/model-routing.md}
- **Model routing rationale:** {1 sentence explaining the routing choice}
- **Generation strategy:** {image_to_video / start_end_image / multi_shot / motion_control}
- **Aspect ratio:** {16:9 / 9:16 / 1:1 / 21:9}
- **Target resolution:** {pixel dimensions from cinematography spec}
- **Resolution parameter:** {720p / 1080p / model-specific equivalent; provider hint
  only until video-generator verifies actual pixels}
- **Model overrides:** {key=value list for live MCP defaults; include audio,
  quality/mode, cfg/guidance, genre, or "none"}
- **Count:** {1 by default; 2 only for review-gated hero/uncertain shots when
  video-generator confirms the live schema supports it}
- **Audio generation:** ambient={on/off}; sfx={on/off}; dialogue={on/off};
  music=off; narration=off; source={generated/none/supplied}

## Frames
- **Start frame:** shots/{shot_id}/start.png
- **End frame:** shots/{shot_id}/end.png
- **Key frames:** {paths or "None"}

## Reference Roles
- **start_image:** {file path}
- **end_image:** {file path}
- **image (subject):** {file path(s)}
- **image (style):** refs/style/style_anchor_01.png

## Reference Audit
- **Required refs:** {character/location/prop/recurring/style refs that must be uploaded}
- **Continuity-critical refs:** {refs named by continuity inventory or Phase 6 report}
- **Baked-frame refs:** {continuity-critical refs that cannot be uploaded to the chosen
  video model and therefore must be present in the start/end frames}
- **Reference priority:** start,end,principal_character,hero_prop,recurring_element,
  location,style
- **Missing or blocked refs:** {None or blocker}

## Prompt

[STYLE] {Global style phrase from keyword library}
[FILMSTOCK] {Filmstock phrase from keyword library}
[SCENE] {Location vocabulary + lighting condition vocabulary + global negative constraints}
[FRAMING] {Frame size, lens, camera mount, motion path}
[PACING] {slow / moderate / fast — with clip-specific meaning}
[ACTION] {transition_description — 2–4 sentences; subject appearance, movement trajectory,
         state changes, existence statements}
[SUBJECT] {Subject key visual features for consistency}
[AUDIO] {Audio direction from Phase 4.5}
[DURATION] {4 / 6 / 8 seconds}

## Audio Generation Preferences
- **Ambient audio:** {on/off}
- **Sound effects:** {on/off}
- **On-screen dialogue/lip-sync:** {on/off; exact lines if on}
- **Music:** off unless diegetic music is visible in-frame
- **Narration:** off; narration is a separate process
- **Audio source:** {generated / none / supplied}
- **Preserve silence:** {true/false}

## Generation Prompt

{Model-native plain text assembled with the algorithm in `references/model-routing.md`.
This is the exact prompt `video-generator` submits to Higgsfield.}

## Consistency Notes
- {Any WARN items from storyboard consistency check}
- {Continuity flags from shot list}
- {Action taken for each WARN/continuity item: regenerated frame, added reference,
  injected prompt constraint, or blocker}

Style Language Rule

Every [STYLE], [FILMSTOCK], and [SCENE] field copies from the project's prompt keyword library. No ad hoc vocabulary.

Continuity Constraint Injection

Before finalizing each prompt, check:

• The location bible negative space rules for this location
• The continuity inventory for this scene
• Any WARN items from Phase 6

Inject all applicable constraints into [SCENE] or [ACTION]. The consistency pass is not informational: action its output now. Fix clear WARN items by regenerating frames or adding references, and convert any remaining WARN/continuity item into an explicit prompt constraint, Reference Audit entry, or blocker.

Manifest

After writing all prompt files, produce:

prompts/manifest.md

# Shot Generation Manifest

| Shot ID | Scene | Duration | Model | Strategy | Aspect | Target Resolution | Resolution Param | Overrides | Count | Audio | Required Refs | Review Gate | Clip Boundary | Transition Type | Transition Duration | Continuity Flags | Start | End | Keys | Prompt File |
|---------|-------|----------|-------|----------|--------|-------------------|------------------|-----------|-------|-------|---------------|-------------|---------------|-----------------|---------------------|------------------|-------|-----|------|-------------|
| S11_SH001 | SC-11 | 8s | seedance_2_0 | start_end_image | 16:9 | 1920x1080 | 1080p | generate_audio=true;quality=standard;genre=auto | 1 | ambient=on;sfx=on;dialogue=off;music=off;narration=off;source=generated | refs/characters/maeve/front.png;refs/props/gannet/primary.png | required | continuous | cut | 0 | eyeline;prop reset | shots/S11_SH001/start.png | shots/S11_SH001/end.png | None | prompts/S11_SH001_prompt.md |

Phase 8: Asset Pipeline

Downstream handoff: Once Phase 8 has produced prompts/manifest.md, prompt files, storyboard frames, model routing, generation strategy, aspect ratio, target resolution, resolution parameter, model overrides, count, required-reference audit, audio-generation preferences, review-gate metadata, and clip-boundary metadata, hand off to video-generator for Higgsfield MCP media upload, job submission, polling, resume handling, and assembly-order output. video-generator is the sole owner of runtime generation logging and Higgsfield job interactions. It must verify required references, apply audio and model overrides, handle resumable job logging, submit and poll jobs, and write assembly order before marking a shot complete. Prerequisite: Read references/asset-pipeline.md.

Maintain a consistent asset pipeline so generated clips can be traced back to their source shots, retaken efficiently, and assembled in the correct order.

File Naming Convention

shots/{shot_id}/start.png          — start frame image
shots/{shot_id}/end.png            — end frame image
shots/{shot_id}/key{NN}.png        — key frame(s)
prompts/{shot_id}_prompt.md        — video generation prompt
generated/{shot_id}/v{N}.mp4       — generated video clip (v1 = first take, v2 = retake)
generated/{shot_id}/selected.mp4   — symlink or copy of the selected take

Runtime Generation Log

Do not write runtime generation logs in this skill. shot-specifier prepares prompts/manifest.md, prompt files, storyboard frames, model routing, aspect ratio, resolution parameters, model overrides, count, required-reference audit, audio-generation preferences, review-gate metadata, and clip-boundary metadata. video-generator owns runtime logging after generation begins.

Reference Files

Related Skills

Templates

Best Practices

One Shot, One Action. Every clip contains one action in one location. If a scene description contains "and then", split it.

Budget First, Decompose Second. Never start writing shot rows without a clip-duration budget. The budget prevents scene collapse.

Direction is Concrete. Actor direction specifies position, action, and eyeline. "Looks concerned" is not direction. "Holds gaze on the empty screen, jaw still, for two seconds then turns toward Iain" is direction.

Storyboards are a Consistency Investment. Generating storyboard frames before video generation catches reference drift early. One regenerated storyboard frame costs far less than one regenerated video clip.

Copy Keyword Library, Do Not Paraphrase. Every [STYLE], [FILMSTOCK], and [SCENE] field copies from the prompt keyword library verbatim. Paraphrasing introduces vocabulary drift.

Inject Negative Constraints on Every Prompt. The location bible's negative space rules (no trees, no US road markings, no blue sky) must appear in every exterior prompt. The model does not remember previous prompts.

Log Every Generation Call. Job IDs are the only way to retrieve or retake a specific clip. Log them immediately. An unlogged job is a lost job.

Retakes Are Part of the Pipeline. Not every first take is the right take. The asset pipeline distinguishes v1 from v2 from selected. Budget for at least one retake per complex shot.

Machine Vision Shots Get Machine Vision Prompts. Drone POV, surveillance feed, and telemetry-screen shots must suppress grain, filmstock, and organic movement. Apply the POV override vocabulary from the keyword library — do not accidentally apply the film-emulation spec to a shot that is supposed to look like a camera drone.