long-video

name	long-video
description	Long video production. Composes a video from a sequence of distinct shots. Trigger ONLY when the user explicitly asks to generate a video longer than 10 seconds (duration > 10s). DO NOT trigger if the requested video is 10 seconds or less.
metadata	{"type":"pattern"}

Trigger condition

Use this pattern ONLY when the user explicitly asks to generate a video longer than 10 seconds (duration > 10s).

Do not use this pattern if:

• The requested video is 10 seconds or less (duration <= 10s). E.g. 4s, 6s, 8s videos. In those cases, use a single video generation step/operation directly.
• The request is for a single animated image or a single clip — use i2v directly instead.

---

Pre-production: scenario document

Before planning any nodes, call plan_text_nodes to emit a scenario document. This externalizes the script breakdown so the user can review and edit it before media generation begins.

The scenario document MUST contain:

1. Visual style anchor — physical description of the world: palette, light quality, surface textures, atmosphere. No mood words — translate everything to observable facts.
2. Shot list — one entry per shot, each with:
   • Scene — location, time of day, physical environment facts.
   • Subject — who or what is in frame and what they do (one verb, one motion arc).
   • Camera — one move, one rhythm word.
   • Duration — 4, 6, or 8 seconds only.
   • Audio — ambient, SFX, music cue, dialogue, or explicit silence.
3. Structure — label each shot by phase:

Phase	Role	Typical duration
Hook	Capture attention — striking visual	4–6s
Build	1–3 scenes that establish context/story	4–6s each
Climax	The emotional or product peak	6–8s
Resolution	Logo, tagline, CTA, or lingering image	4–6s

Use format: "scenario" and a descriptive title (e.g. "Lumino — Trailer Architecture").

After emitting the scenario, proceed directly to plan_production. Reference each shot's scene description in the corresponding promptIntent fields — do not re-derive them from the original user message.

If a canvas text node with format: scenario already exists, read its content from the canvas context and use it as the grounding source. Do NOT emit a new scenario.

Adjust shot count to fit the brief: a 15-second ad needs 3–4 shots, a 30-second film needs 5–7.

---

Mandatory node order — no exceptions

Every shot in a multi-shot production follows this fixed sequence:

Phase 0: t2i reference nodes   (characters, hero props, settings — generated first, in parallel)
Phase 1: t2i keyframe          (one per shot — always generated, never skipped)
Phase 2: i2v animation         (depends_on the keyframe from Phase 1 — always)
Phase 3: concat                (assembles all i2v nodes — always last)

t2v is forbidden in multi-shot production. Use t2i → i2v for every shot, always. t2v skips the keyframe and produces subjects that drift across shots.

Existing canvas nodes are forbidden as direct input to i2v. A product photo, reference image, or any node already on the canvas must never be wired directly into an i2v. It may be used as subject_ref or style_ref on a t2i keyframe — but generating the keyframe is mandatory regardless.

---

Phase 0: Reference generation

Generate these before any shot keyframe. They anchor visual identity across shots. Skip none that apply.

Character reference (one per human subject appearing in more than one shot):

operation: t2i
promptIntent: "Full-body neutral-pose reference of [description]. Front-facing, clean background, even lighting. No action, no set."
label: "Ref — [Name / role]"

Hero prop reference (one per object that must look consistent across shots — product, vehicle, garment, tool):

operation: t2i
promptIntent: "Isolated [prop description] on neutral background, lit to reveal shape, texture, and material. No hands. No scene."
label: "Ref — [Prop name]"

Setting reference (one per location used in more than one shot):

operation: t2i
promptIntent: "Wide establishing shot of [location]. [Physical facts: architecture, surfaces, light quality, time of day]. No people."
label: "Ref — [Location]"

These nodes have no depends_on edges. They run in parallel.

---

Phase 1–2: Shot keyframe + animation (repeat per shot)

For every shot, generate the keyframe first, then animate it:

t2i keyframe:
  operation: t2i
  promptIntent: "[Scene physical facts]. [Subject pose and position in frame]. [Lens, focal length, lighting facts]."
  edges:
    subject_ref: character or hero prop reference (if applicable)
    style_ref: setting reference (if applicable)
  label: "Shot N — [descriptor] keyframe"

i2v animation:
  operation: i2v
  promptIntent: "[ACTION: one verb, one motion arc]. [CAMERA: one move + rhythm word]. [AUDIO: ambient / SFX / music / dialogue / explicit silence]. Duration: [4|6|8]s."
  edges:
    depends_on: the t2i keyframe above — mandatory
    subject_ref: character or hero prop reference (if applicable)
  duration: 4 | 6 | 8
  label: "Shot N — animated"

---

Phase 3: Concat

concat:
  operation: concat
  promptIntent: "Assemble shots in narrative order. No transition effects unless user specified."
  edges: depends_on from each i2v node, in sequence order
  label: "Final cut"

---

Edge rules

• i2v depends on its t2i keyframe. Always. No direct wire from any canvas node or reference to i2v.
• Character/prop references → subject_ref on every t2i and i2v featuring them.
• Setting references → style_ref on every t2i set in that location.
• concat depends on all i2v nodes in shot order. Never depends on t2i nodes directly.

---

Model selection

Always use the canvas default model for every node unless the user explicitly names a model in their request. Do not upgrade to veo-3.1-generate-001 because the output is "an ad" or "final quality" — that is the user's decision to make, not yours.

---

Pacing rules

• Keep each i2v at 4s unless the shot has complex action or camera movement — then use 6s or 8s.
• The total assembled duration is the sum of all i2v durations. Check it against the user's stated length before committing.
• A 30-second spot = roughly 5–7 shots averaging 5s each. A 15-second ad = 3–4 shots.
• Resist stacking every shot at 8s — the edit will feel slow. Vary rhythm: short punchy cuts (4s) followed by a longer breathing shot (6–8s).

---

Audio continuity

Treat audio as a separate layer with its own continuity:

• NO PARALLEL AUDIO NODES: The concat primitive only joins video clips; it does not support overlaying or mixing separate audio tracks. Never plan separate t2s or t2m nodes and connect them to video or concat nodes.
• WRITE AUDIO IN VIDEO PROMPTS: Instead, specify all shot-specific dialogue, music, or room tone directly in the individual shot (i2v) prompts' [AUDIO] section. The video model will generate the audio track natively within each clip.
• If music and ambient coexist, note the balance: "Coffee shop murmur at -28dB beneath slow piano at -18dB."

---

Worked example — 2-shot UGC product ad, 8s, 9:16

Brief: "Make a short UGC ad for this skincare product, 8 seconds, 9:16." User has a product photo on the canvas.

Script breakdown:

• Shot 1 (4s): Creator in bathroom, holds product up, speaks to camera.
• Shot 2 (4s): Close-up of glowing skin the morning after, natural light.

Correct plan — 6 nodes total:

Node 1 — Phase 0: product reference
  operation: t2i
  promptIntent: "Isolated [product name] tube on a clean white background. Front-facing, even studio lighting, sharp focus on label and packaging texture. No hands. No scene."
  label: "Ref — Product"
  edges: none

Node 2 — Phase 0: character reference
  operation: t2i
  promptIntent: "Full-body neutral-pose reference of young woman, late 20s, natural makeup, light hair. Front-facing, clean white background, soft even lighting. No action, no props."
  label: "Ref — Creator"
  edges: none

Node 3 — Phase 1: Shot 1 keyframe
  operation: t2i
  promptIntent: "Young woman in white-tiled bathroom, holding skincare tube up at chest height toward camera. Slight forward lean. Warm ring-light directly behind camera. Shallow depth of field, background softly defocused. 9:16 vertical frame."
  edges:
    subject_ref: Node 1 (product ref)
    subject_ref: Node 2 (character ref)
  label: "Shot 1 — bathroom intro keyframe"

Node 4 — Phase 2: Shot 1 animation
  operation: i2v
  promptIntent: "ACTION: She rotates the product slightly toward lens, eyebrows raise, lips part into a wide smile. CAMERA: Gentle handheld sway, slow push-in. AUDIO: Dialogue — 'You guys, I am obsessed with this.' Room tone at -30dB. No music. Duration: 4s."
  edges:
    depends_on: Node 3 (Shot 1 keyframe)
    subject_ref: Node 2 (character ref)
  duration: 4
  label: "Shot 1 — animated"

Node 5 — Phase 1: Shot 2 keyframe
  operation: t2i
  promptIntent: "Close-up of young woman's face, right cheek in sharp focus. Morning light from window at 45-degree angle from upper left. Visible skin texture, soft natural glow. No product in frame. 9:16 vertical."
  edges:
    subject_ref: Node 2 (character ref)
  label: "Shot 2 — morning skin keyframe"

Node 6 — Phase 2: Shot 2 animation
  operation: i2v
  promptIntent: "ACTION: She presses fingertips gently to cheekbone, skin rebounds to show elasticity. CAMERA: Organic handheld sway. AUDIO: Soft morning birdsong at low volume, no dialogue, no music. Duration: 4s."
  edges:
    depends_on: Node 5 (Shot 2 keyframe)
    subject_ref: Node 2 (character ref)
  duration: 4
  label: "Shot 2 — animated"

Node 7 — Phase 3: Concat
  operation: concat
  promptIntent: "Assemble Shot 1 then Shot 2 in order. No transitions."
  edges:
    depends_on: Node 4, Node 6
  label: "Final cut"

What is absent from this plan and why:

• No t2v nodes. Every video starts from a t2i keyframe.
• The product photo from the canvas does not appear as a direct edge into any i2v. It informed Node 1 (the product reference), which then flows through the keyframes.
• The character does not exist on the canvas — a reference was generated from scratch.

---

Common failures

• Using t2v for any shot. Forbidden. Always use t2i → i2v.
• Wiring a canvas node or reference image directly into i2v. Forbidden. Generate a t2i keyframe first. The canvas node can be subject_ref on the keyframe — but the keyframe is never optional.
• Skipping Phase 0 references. Skipping character references causes identity drift across shots. Skipping product/hero prop references causes inconsistent object appearance. Skipping setting references causes location incoherence. Generate all that apply, every time.
• Skipping the keyframe because "the reference is good enough". It is not. The reference defines identity; the keyframe defines framing, lighting, and pose for that specific shot. They are different nodes with different purposes.
• Wiring t2i nodes directly into concat. Produces a slideshow of stills. Always animate first.
• Mood words without physical grounding. "Cinematic", "epic", "emotional" produce generic output. Translate to physical facts before passing to PromptEngineer.
• All shots at maximum duration. Vary rhythm — short punchy cuts (4s) followed by longer breathing shots (6–8s).
• Omitting audio in any i2v node. Always specify, even if "No dialogue. No music. Room tone only."