---
name: gta-gameplay-ad
description: "Your product becomes a power-up in a fake GTA-style game: the HUD reacts, the hero completes the mission. Just paste your product link."
---

# GTA gameplay ad (Nano Banana 2.1 + Seedance 2.5)

Source: public Arcads workflow `flow/public/8saanuazxa81OxaFauVMAcQf`, with every original prompt kept word for word in the section "refs/original-workflow-prompts.md" at the end of this file. This skill keeps its creative engine:
- research
- master brief
- the product as a game mechanic
- the prompt rules for HUD, cash and first frames

It changes three things:
1. **The length is free.** The angle decides it (section 4); the original was always 10 × 3 s.
2. **Few video generations.** Seedance 2.5 makes up to 30 s per clip with hard cuts, one voice and one character. One clip covers a whole block of shots: 1 clip for 15 s, 2 for 30 s, 2–3 for 60 s. The original needed 10 Kling clips.
3. **Better images.** Nano Banana 2.1 at 2K with `thinkingLevel: "HIGH"` handles HUD text, logos and character consistency (up to 14 references).

**Status: untested live.** Write every lesson from the first real run back into this file.

Follow the steps in order and stop at every ✋ for the user's approval. Raise problems instead of improvising around them.

## 1. Non-negotiable rules

1. **Models:**
   - Images: `arcads_generate_image_nano_banana`, `model: "nano-banana-2-1"`, `resolution: "2K"`, `thinkingLevel: "HIGH"`.
   - Video: `arcads_generate_video_seedance_25`, `audioEnabled: true`, `resolution: "1080p"` (720p if the user wants it cheaper).
   - Always `nbGenerations: 1`. No test renders, no extra inserts.
2. **Fewest clips.** Group shots into clips of **≤15 s by default**. Go up to 30 s per clip only if the user asks for the fewest generations; it is untested whether quality holds over 30 s. The photoreal hero close is the **last shot of the last clip** (a hard cut out of the game look), not a separate clip.
3. **Project:** call `arcads_create_project` before the first generation and pass `projectId` on every call. Give every asset a readable `title` (`<brand>_gta_clipA_v1`).
4. **Format:** ask once. The default is **9:16** for social. In 9:16, the HUD sits inside the safe zone: cash and wanted stars top-right below the top 10 %, minimap bottom-left above the bottom 20 % (captions go there). For 16:9, use the original layout.
5. **Style only, no IP.**
   - Prompts may say "GTA V gameplay" as a style word.
   - The frame never shows the words GTA, Grand Theft Auto, Rockstar or Los Santos, a franchise logo, or a real game character.
   - The title card is the **brand name** in a blocky heist-game font.
   - If the content filter blocks a prompt, write "open-world crime video game, 2013 AAA console look" instead, and say so.
6. **No real people, no named rivals on screen.** The research can name rival brands and figures. On screen, the rival's product is generic and unbranded, and real people only shape the VO delivery.
7. **The product appears only in shots that carry it** (from the equip/reveal shot onwards, never in rival or peer cutaways, never as a stray prop).
8. **Seedance adds music even when told not to** (seen on LegalClaim). Every clip prompt still says "no music". If the review hears music, run `arcads_isolate_voice` on the stitched cut before laying our bed, and tell the user the game SFX go with it.
9. **Pronunciation:** run every spoken line through `seedance-phonetic-pronunciation` (brand names, ingredients, numbers). Captions keep the real spelling.

## 2. Step 0: inputs (one message)

Ask for:
- product photo(s)
- product URL
- logo (PNG)
- format (9:16 or 16:9)
- VO language
- claims the client forbids
- a target length, if they have one (otherwise the angle picks it)

Create the work folder `GTA/<Brand>_<slug>/plan.md` (template in the appendix). Every asset ID, credit cost and review verdict goes there.

## 3. Step 1: research and brief (Claude does the "LLM nodes")

Do the work of the original's Gemini nodes yourself, following the original prompts in the section "refs/original-workflow-prompts.md" at the end of this file:

1. **Product Analysis** (refs section 1). Research the URL and the web (WebSearch / firecrawl) and fill the 5-layer JSON: product, audience, the category right now, culture and rivals, pain points and 3–5 angles. Be concrete: names, numbers, current memes.
2. **Master Brief** (refs section 2), all 7 steps:
   - one locked protagonist, plus rival and peer
   - satire target and fake headline
   - the audience's real fantasy
   - product-in-world signatures (plate, title card, app icon, flex placements)
   - **the game mechanic** (benefit → HUD bar or power-up)
   - before → break → after spine
   - guardrails and hero close
   - neutral, HUD-free `character_description`
3. **Pick the length and format from the angle** (section 4) and say why in one line.
4. **Shotlist**: use the Creative Brief rules (refs section 3) on the chosen format's skeleton:
   - every gameplay shot has an action verb
   - picture and line describe the same moment
   - the VO lines read as one script ending on the CTA
   - the peer talks *about* the protagonist
   - one VO profile
   - **at most 9 words per 3 s shot** (Seedance speaks ~3 words/s and needs air before the cut)

✋ **Show the user:**
- the angle and length, each in one sentence
- the protagonist, rival and peer
- the game mechanic
- the shotlist table (shot, camera, what we see, line), grouped by clip

Also list the claims to validate. Wait for a yes and apply their edits.

## 4. Formats: the angle decides the length

Each shot is ~3 s. The beats in **bold** are always there; the others are added as the length grows.

| Format | Length | Clips | Beats | Good for |
|---|---|---|---|---|
| **Punch** | 15 s | 1 × 15 s | **Status quo (before)** → **Equip** (mechanic triggers) → **Title card** → **Flex** (after) → **Hero + CTA** | One sharp benefit, hook-test variants, retargeting |
| **Classic** | 30 s | 2 × 15 s | The original 10 shots: status quo, world, rival satire, equip, title card, flex + drink, flex B, peer hype, action, hero + CTA | Default; most angles |
| **Mission** | 45–60 s | 3 × 15 s, or 2 × 30 s if the fewest clips is asked | A mission: "MISSION: [goal]" objective card → before + rival → equip → 2–3 gameplay set pieces (chase, heist, race) each clearing an objective → "MISSION PASSED" + cash → peer hype → hero + CTA | Story-rich angles, launches, products with several benefits |

Rules that hold at every length:
- the product first appears at the equip, at ~1/3 of the runtime
- the brand title card comes right after the equip
- the cash climbs across gameplay shots, with a jump at the equip
- the HUD appears only on gameplay shots, never on cutscenes, the title card or the hero
- the hero close is photoreal for a physical product (game render for an app or service) and is placed in a story setting, never a void

## 5. Step 2: look references (Nano Banana 2.1)

1. **Style reference**: the 2×2 grid prompt from refs section 0, in the chosen ratio.
2. **Protagonist sheet**: the prompt per refs section 5 (turnaround, close-up plus full body, **neutral baseline**, **no HUD**). Reference: the style ref.
3. **Rival / peer sheets**, only if the format has them. Use the same 16:9 sheet layout as the protagonist (front, side, back, close-up), no HUD. **Reference: the approved protagonist sheet, not the style ref**, with "EXACTLY the same render style, lighting, background and layout as the attached sheet, same game engine, same artist; do NOT copy the person". (eSIM io run: stills referenced only on the style grid came out in other styles, one illustrated, one more realistic. The user asked for them to match the protagonist, and the sheet reference fixed it on the first try.)

✋ Show them together and wait for approval. Regenerate only what is rejected.

## 6. Step 3: keyframes (Nano Banana 2.1)

The keyframes lock what Seedance gets wrong when it works only from text: HUD text, the cash figure, the title card, the product label. Make one keyframe for:
- the **first shot of each clip**
- every shot with **on-screen text or HUD changes**
- every shot showing the **product**
- the **title card**
- the **hero**

Other shots run from the prompt plus the sheets.

Write each one with the image rules of the Prompt Writer (refs section 4):
- a calm **first frame**: no motion blur, flashes, equip pop-ups or HUD mid-animation
- HUD with the exact cash figure in quotes, only on gameplay shots
- mechanic bar low before, full after
- protagonist prompts say "the protagonist from the attached character sheet, same face, hair, build and exact outfit, NOT different clothes", plus the BEFORE/AFTER condition

References: style ref plus the sheets of whoever is in frame, plus product photo(s) for product shots, plus the logo for the title card and hero. The hero uses no style ref, because it is photoreal.

Run them in parallel. Check each one yourself (`arcads_view_frame`): HUD placement, rising cash, readable text, correct label, same face. ✋ Show them as a storyboard (in shot order, with the line under each) and wait for approval. Fix only the rejected ones.

## 7. Step 4: clips (Seedance 2.5, all clips in parallel)

One call per clip.

**References:**
- that clip's keyframes, in shot order
- the protagonist sheet
- the rival/peer still if they are in the clip
- the product photo if the product is in the clip
- the logo if the clip has the title card or hero

Number them in the prompt.

**Prompt template:**

```
Fast-cut 3D video game ad in GTA V gameplay look: glossy semi-realistic game-mesh render,
in-game camera grammar, saturated game colour grade, never live-action film
[except the final photoreal hero shot].
References: @Image1 = first frame of shot 1, @Image2 = first frame of shot 3, ... (each listed shot opens
exactly on its image and keeps its framing, HUD layout and on-screen text). @ImageN = protagonist identity
sheet: same face, hair and outfit in every shot. [@Image rival / peer / product / logo.]
Narrator VO: [vo_profile]. Same narrator voice through the whole clip.

[0:00-0:03] Shot 1 — [one action + one camera move]. VO: "[line]"
[0:03-0:06] Shot 2 — [...]. HUD: the [MECHANIC] bar flashes red and empty, cash "$1,240". VO: "[line]"
[0:06-0:09] Shot 3 — [rival action]. The rival says, lip-synced: "[line]"
[0:09-0:12] Shot 4 — smash-zoom as he equips the [product], [signature flash], the bar surges to full,
cash jumps to "$45,000", he snaps from drained to switched on. VO: "[line]"
[0:12-0:15] Shot 5 — aerial sweep over a sunset city, title card "[BRAND]" holds centred. VO: "[line]"

Hard cut between every shot. Each line starts within 0.3 s of its shot and ends before the cut.
Game SFX allowed (HUD chimes, whooshes, the product's own sound). No music, no score.
```

The last clip ends on: `[...] Shot N — HARD CUT out of the game: photoreal real-world product shot, slow push-in on [product] in [story setting], logo small, "[url]" on screen. VO: "[CTA]"`.

**Same narrator across clips:**
- Clips run in parallel, so each prompt repeats the same `vo_profile` word for word.
- If the voices differ at review, extract ~6 s of clean VO from the best clip (`arcads_extract_audio` → `arcads_isolate_voice`).
- Then regenerate only the odd clip with it as `referenceAudios` and "Narrator voice = @Audio1".

Character lines are lip-synced inside the clip. Never use audio_driven or omnihuman.

**Review every clip yourself** (`arcads_view_contact_sheet` + `arcads_view_frame` at each cut, `arcads_analyze_media` for audio):
- each listed shot opens on its keyframe
- same protagonist
- HUD only on gameplay shots, cash rising
- product only where planned
- every line complete and correctly pronounced
- same narrator in every clip
- music present or not

Regenerate a clip only for a failed check, and tell the user which one.

## 8. Step 5: music and edit

1. **Music**: `arcads_create_music`, model `elevenlabs`, `instrumentalOnly: true`, `sections` adding up to the exact final length, with boundaries on the cuts. Use musical terms only, no game, artist or franchise names (the provider rejects them). Shape:
   - **Fog** until the equip: filtered sub drone, glitchy hats, anxious
   - **Build**: 2–3 s riser, near-silent last half-second
   - **Drop** on the equip / title card: punchy drums, driving bass, arpeggios
   - **Peak** on the biggest set piece
   - **Resolve** on the hero: drums stop dead, airy pad, one clean digital ping

   Pick the genre from the audience research (phonk, trap, drill, synthwave, West Coast bounce…).
2. **Stitch** the clips (`arcads_stitch_videos`). If music leaked, run `arcads_isolate_voice` first (rule 8).
3. **Layer** the music (`arcads_layer_videos`, music as an `audio` layer at ~0.16 under the voice).
4. **Captions** only if asked (`arcads_add_subtitles`, bottom safe zone, real spelling).
5. Check the final with `arcads_probe_media`. Show the user **only the final**, never the parts.

## Appendix: plan.md template

```
# GTA gameplay ad — <Brand> (<format>, <length>, <ratio>, SD2.5 <res>, <lang>)
## Inputs: product photos · URL · logo · forbidden claims
## Research summary (5 layers, short) · Chosen angle · Why this length
## Master brief: protagonist · rival · peer · mechanic · spine · product-in-world · satire headline · guardrails
## Shotlist grouped by clip (shot / beat / camera / state / product / keyframe? / line / words)
## Claims to validate
## Assets: style ref · sheets · keyframes · clips · music · final (IDs, credits)
## Review log (per clip: checks, verdict) · Lessons for the skill
```

# refs/original-workflow-prompts.md

# GTA 6 Style Ad — prompts extraits du workflow Arcads public 8saanuazxa81OxaFauVMAcQf

## 0. STYLE REFERENCE IMAGE — prompt (Nano Banana, 16:9)

```
A 16:9 four-panel 2x2 grid, each quadrant a separate frame in a GTA V gameplay
aesthetic — glossy semi-realistic 3D game-mesh character renders with a slightly
blocky, video-game look (not photoreal film), oversaturated game colour grade,
each panel showing the on-screen GTA game HUD: cash counter top-right, five-star
wanted rating, minimap bottom-left, location banner bottom-right.

Top-left: third-person over-the-shoulder gameplay shot from BEHIND a young man
in a messy gamer apartment, seen from the back of his head and shoulders like a
GTA player camera, cool RGB/blue interior lighting, full HUD overlay.

Top-right: first-person FPV driving shot from inside a car down a palm-lined
Los Santos street at golden hour, hands on the wheel, sun flare, full HUD overlay.

Bottom-left: in-game cutscene of two glossy 3D character NPCs talking in a
scene — game-render skin and lighting, mid-shot, clearly rendered game footage,
HUD overlay.

Bottom-right: third-person gameplay shot of a character standing in a room
holding a drink can like an equipped in-game item, glossy game-mesh render,
HUD overlay.

Clean thin black gutters between panels. The whole image must look like four real
screenshots from Grand Theft Auto V gameplay — blocky game render, not cinematic
photography, not film.
```

## 1. PRODUCT ANALYSIS — system prompt (Gemini 3.1 Pro, input: product image + product/URL)

```
You are an elite advertising strategist and cultural researcher specialising in
internet-native, meme-literate marketing. You research the way a terminally-online
performance marketer does — you know the rival brands, the in-jokes, the recent
drama, and the figures every niche follows.

You have been given:
- A product image
- A product URL

Your job is to build a deep intelligence brief on this product and the world it
lives in. This brief feeds a downstream creative node that will build a chaotic,
GTA-gameplay-style hype ad full of parody, cultural references, and competitor
jabs. The richer and more current your research, the sharper that ad becomes.
Surface raw material — do not make the creative decisions yourself. Your job is
to find everything; the next node chooses what to use.

Research in five layers. Be specific and concrete at every layer — names, numbers,
real references. Generic observations are useless downstream.

LAYER 1 — THE PRODUCT
Read the URL and the image. Extract:
- Name, category, price point, what it actually does
- Key claims and selling points (verbatim where strong)
- Form factor, colours, packaging, anything visually distinctive
- What plan/tier/format a first-time buyer would land on

LAYER 2 — THE AUDIENCE
Profile who actually buys this:
- Age, gender skew, income, where they spend time online
- Their internet subculture — what communities, platforms, creators they follow
- What they aspire to, and what status signals matter to them
- The fantasy this product is adjacent to (wealth, freedom, status, belonging)

LAYER 3 — THE CATEGORY RIGHT NOW
Research what is happening in this space currently:
- Recent news, launches, controversies, or shifts in the last few months
- The dominant trend or narrative in the category right now
- Anything timely the ad could ride or reference to feel current

LAYER 4 — THE CULTURE & RIVALS
This is the raw material for parody and jabs. Find:
- Direct competitors and rival brands (named) — especially ones with a mockable
  reputation or aesthetic
- The memes, catchphrases, and running jokes native to this audience
- Recognisable figures, archetypes, or personalities this audience knows
  (the "guru", the rival founder, the famous customer type)
- Any ad, format, or cultural moment in this niche that's been parodied before

LAYER 5 — PAIN POINTS & ANGLES
- The core frustrations this product removes, in the audience's own blunt voice
- A shortlist of 3–5 distinct angles the ad could take — each a different
  emotional or comedic entry point, not variations of one idea

Output strict JSON in exactly this shape. No commentary before or after.

{
  "product": {
    "name": "",
    "category": "",
    "price_point": "",
    "what_it_does": "",
    "key_claims": [],
    "visual_identity": ""
  },
  "audience": {
    "demographic": "",
    "subculture": "",
    "aspirations": "",
    "adjacent_fantasy": ""
  },
  "category_now": {
    "recent_developments": [],
    "dominant_narrative": "",
    "timely_hooks": []
  },
  "culture": {
    "rival_brands": [],
    "memes_and_catchphrases": [],
    "recognisable_figures": [],
    "parody_targets": []
  },
  "pain_points": [],
  "candidate_angles": [
    { "angle": "", "entry_point": "" }
  ]
}
```

## 2. MASTER BRIEF — system prompt (Gemini 3.1 Pro, input: product analysis + product image)

```
You are an elite creative director for internet-native, meme-literate ad
campaigns. You make viral hype ads in the GTA-gameplay-parody style — the format
that sells a product through a chaotic, satirical, fake-video-game world built
from the audience's own culture, in-jokes, and references. You are terminally
online and fluent in whatever subculture the product belongs to.
You have been given:
A research JSON brief on the product, its audience, the category, and the
culture around it, including a shortlist of candidate angles.
Your job is to make the creative decisions for a 30-second GTA-gameplay-style
hype ad and output a locked creative brief. Downstream nodes turn your brief into
a 10-shot script and then into generation prompts, and they follow your decisions
exactly — so commit fully. Do not hand down options. Choose.
Everything you decide must come from THIS product's research — its real audience,
its real culture, its real rivals. The GTA-gameplay style is the only fixed
ingredient; every other choice (the tone, the fantasy, the targets, the
references, the character) is dictated by the research, not by genre defaults. If
a choice would fit any product, it is too generic — push until it could only fit
this one.
Work through the following internal steps before writing your output. Do not show
your working — only output the final JSON.
STEP 1 — INVENT AND LOCK ONE CONCRETE PROTAGONIST
This is the single recurring character the audience follows through the whole ad.
You are NOT writing a style or demographic "type" — you are casting one specific,
named individual who will be rendered as the SAME person in every shot they
appear in.
Their gender, age, ethnicity, and identity come FROM THE RESEARCH — who this
audience sees as "them" or as the person they want to become. Never default to
a particular gender; let the audience profile decide.
Lock concrete, visible details: name, age, face (shape, hair, facial hair if
any, distinctive features), build, exact wardrobe, and the energy they carry.
Specific beats generic. "Mid-20s, lean, buzzcut, oversized vintage tee, tired
eyes that sharpen later" beats "a young person who looks relatable."
STEP 2 — LOCK THE SUPPORTING CAST
Two concrete supporting characters, each cast as specifically as the protagonist:
THE RIVAL (speaks in shot 3) — the face of the satire target. The person the
audience already loves to mock.
THE PEER (speaks in shot 8) — a character in the protagonist's world who REACTS
to the protagonist's transformation, hyping it up enthusiastically to someone
else (the protagonist is off-screen in that shot). NOT a random person
independently enjoying the product — their line is about OUR protagonist's change.
Give each a name and concrete look. They are never the protagonist.
STEP 3 — RUN THE PROVOCATIONS
Your first instinct for each will be the obvious one everyone has seen. Reject
it. Push to the specific version pulled from the research. For each, decide and
commit:
PROVOCATION A — THE SATIRE TARGET
What does this ad mock? The format thrives on satirising something the audience
already has feelings about — a rival brand, a category cliché, a recognisable
figure, an outdated "normal" way of doing things. Pick the target from the
research, then write a fake breaking-news headline that mocks it. Specific
enough that this audience laughs in recognition. Generic = failure.
PROVOCATION B — THE ASPIRATIONAL FANTASY
Every audience fantasises about a different version of "made it." Name the
SPECIFIC fantasy world this audience actually holds — the settings, objects,
and status signals that mean success to THEM. Do not default to cars and money
unless the research shows that is genuinely this audience's fantasy.
PROVOCATION C — THE PRODUCT IN THE WORLD
How does THIS product physically live inside the GTA-style world so it feels
native, not pasted on? Decide its in-world signatures: a vehicle licence plate,
the title-card brand name in the blocky game font, an in-game phone app icon,
and how the product appears across the flex/hero shots. Make every signature
specific to this product's name and identity.
STEP 4 — TRANSLATE THE BENEFIT INTO A GAME MECHANIC (required)
This is what makes the product feel native to GTA instead of an ad pasted into a
game. Take the product's REAL benefit and express it as an in-game mechanic the
player would recognise — a HUD bar that fills, a stat that boosts, a power-up the
character equips, a status effect, an ability unlock.
Example shape: an energy drink → drinking the can fills a "STAMINA" bar and the
screen surges with speed lines; a focus product → a "FOCUS" / "COGNITIVE
CAPACITY" bar replaces the GTA health/armor HUD and the blurry world snaps
sharp on consumption.
You MUST invent one even when the product has no obvious power-up (a course, an
app, a service). Push to find the mapping — a skill tree that unlocks, a level
that loads, a "WANTED" rating that drops, an ability the character gains.
Name the mechanic, say which GTA HUD element it lives in / replaces, and
describe the on-screen change when the product is consumed/used.
STEP 5 — WRITE THE MESSAGE SPINE
This is the story arc, not final VO copy. ONE character's journey: BEFORE → BREAK
→ AFTER, told THROUGH the game mechanic from STEP 4.
BEFORE: the protagonist living the wrong way — show it as a game state (low
bar, glitching/lagging world, chaos, a high "WANTED" level for the old way).
BREAK: they find/equip the product — the mechanic from STEP 4 triggers.
AFTER: the new reality — the bar fills, the world sharpens, the character is
transformed and in control.
Every beat must work on TWO levels at once: what the audience really feels, AND
how that reads as gameplay. This crossover is the whole point — keep them fused.
STEP 6 — SET GUARDRAILS AND THE HERO CLOSE
Decide the tone guardrails (what to lean into and what to avoid for this specific
brand and audience) and how the final beat works — the hard cut OUT of the game
world into the real, photoreal product hero shot using the actual product image.
STEP 7 — WRITE THE CHARACTER DESCRIPTION (for the reference-sheet node)
Render a single, self-contained description of the locked protagonist from STEP 1,
written so a downstream image node can generate a GTA V character REFERENCE SHEET
from it. This text feeds an image generator that also receives the fixed Style
Reference photo, so:
Describe the protagonist as a GTA V gameplay-style 3D game-mesh character —
glossy semi-realistic game render, in-engine skin and lighting — matching the
Style Reference look. NOT photoreal, NOT a real photo of a person.
Be concrete and exhaustive on everything that must stay consistent: face shape,
hair style, build, exact wardrobe, colours, accessories. This is the single
source of truth for the character's IDENTITY in every later shot.
CRITICAL — DESCRIBE A NEUTRAL BASELINE, NOT A "BEFORE" STATE. The reference
sheet captures WHO the character is, not how their life is going. Do NOT bake in
the before-state: no tired/hollow eyes, no dark under-eye circles, no slumped
posture, no greasy or messy hair, no rumpled/sweaty clothes, no gloomy mood.
Render them neutral, even, well-lit, composed — a clean baseline. The downstream
shots twist this baseline into "before" (dishevelled, drained) or "after"
(sharp, transformed); if the baseline is already dishevelled, the after-shots
inherit it and the transformation never lands.
Keep it audience-appropriate per the research.
NO GAME HUD ON THE SHEET. The reference sheet must be a clean character
turnaround ONLY — absolutely no GTA HUD elements: no cash counter, no minimap, no
wanted stars, no stat bars, no on-screen text. (A HUD baked into the sheet gets
copied into every shot that uses it as a reference — e.g. a fixed cash figure —
which freezes the climbing cash counter downstream. Keep the sheet HUD-free.)
Write it as a clean descriptive block ready to drop into an image prompt — no
shot framing, no multi-view instructions (the sheet node handles layout).
Output strict JSON in exactly this shape. No commentary before or after.
{
"chosen_angle": "",
"message_spine": {
"before": "",
"break": "",
"after": ""
},
"game_mechanic": {
"name": "",
"hud_element": "",
"on_use_change": ""
},
"protagonist": {
"name": "",
"age": "",
"identity": "",
"face": "",
"build": "",
"wardrobe": "",
"energy": ""
},
"supporting_cast": {
"rival": { "name": "", "description": "" },
"peer": { "name": "", "description": "" }
},
"product_in_world": {
"plate": "",
"title_card_name": "",
"app_icon_treatment": "",
"flex_placements": ""
},
"satire_beat": {
"target": "",
"fake_headline": ""
},
"aspirational_fantasy": "",
"tone_guardrails": "",
"hero_close": "",
"character_description": ""
}
```

## 3. CREATIVE BRIEF — system prompt (Gemini 3.1 Pro, input: master brief + product image + logo)

```
You are a world-class creative director for internet-native, GTA-gameplay-style
video ads. You write the SCRIPT and SHOTLIST — the creative blueprint — for a
30-second ad. A separate downstream node turns your shotlist into generation
prompts; your job is purely the creative and narrative structure, output as clean
structured data.

You have been given:
- A locked creative brief JSON: chosen_angle, message_spine (before / break /
  after), game_mechanic, protagonist (locked, named individual), supporting_cast
  (rival, peer), product_in_world signatures, satire_beat, aspirational_fantasy,
  tone_guardrails, hero_close, and character_description.

THE LOCKED CAST — USE THEM, DON'T REINVENT THEM
The protagonist and supporting cast are already cast as specific people in the
brief. You do not redesign their looks; a separate image node builds a character
reference sheet from the brief's character_description, and every shot will render
that same person. Your job is to say WHO is on screen and WHAT they do — refer to
them as "the protagonist", "the rival", "the peer", never as a new or different
person.

THE FIXED SKELETON — DO NOT CHANGE
The ad is always 10 shots, ~2–3s each. The per-shot FLAGS below are FIXED and
identical for every product. You must NOT move, add, or remove them. You only
write the CONTENT (visual concept and spoken line) for each shot. The flags tell
you each shot's role, camera, who is on screen, the protagonist's state, whether
it carries the product and logo, and what kind of audio it has.

Shot 1  — STATUS QUO        | cam: CUTSCENE     | who: PROTAGONIST solo        | state: BEFORE     | product: NO  | logo: NO  | audio: VO
Shot 2  — WORLD ESTABLISH   | cam: FPV/BEHIND   | who: PROTAGONIST solo        | state: BEFORE     | product: NO  | logo: NO  | audio: VO
Shot 3  — RIVAL / SATIRE    | cam: CUTSCENE     | who: RIVAL (no protagonist)  | state: NONE       | product: NO  | logo: NO  | audio: CHARACTER
Shot 4  — THE BREAK / EQUIP | cam: FPV/BEHIND   | who: PROTAGONIST solo        | state: TRANSITION | product: YES | logo: NO  | audio: VO
Shot 5  — TITLE CARD        | cam: CUTSCENE     | who: NO PEOPLE               | state: NONE       | product: NO  | logo: NO  | audio: VO
Shot 6  — FLEX + DRINK      | cam: FPV/BEHIND   | who: PROTAGONIST solo        | state: AFTER      | product: YES | logo: NO  | audio: VO
Shot 7  — FLEX B + PRODUCT  | cam: THIRD-PERSON | who: PROTAGONIST solo        | state: AFTER      | product: YES | logo: NO  | audio: VO
Shot 8  — PEER HYPE (COMIC) | cam: CUTSCENE     | who: PEER (no protagonist)   | state: NONE       | product: NO  | logo: NO  | audio: CHARACTER
Shot 9  — ACTION + PRODUCT  | cam: FPV/BEHIND   | who: PROTAGONIST solo        | state: AFTER      | product: YES | logo: NO  | audio: VO
Shot 10 — HERO CLOSE + CTA  | cam: HERO CLOSE   | who: NO PEOPLE               | state: NONE       | product: YES | logo: YES | audio: VO (CTA)

WHO-IS-ON-SCREEN RULES (FIXED):
- "PROTAGONIST solo" → only the locked protagonist is on screen.
- "RIVAL (no protagonist)" → shot 3 is the satire target alone; the protagonist
  must NOT appear, and the product must NOT appear.
- "PEER (no protagonist)" → shot 8 is the PEER alone (the protagonist is OFF-screen)
  — see the shot 8 rule below. No product on screen.
- "NO PEOPLE" → shots 5 and 10 have no characters (shot 10 still carries a VO CTA
  over the product, just no one on screen).

PROTAGONIST STATE (BEFORE / AFTER) — FIXED:
The character reference sheet is a NEUTRAL baseline of the protagonist (just who
they are). Each shot twists that baseline into a life-state. This is what sells the
transformation, so honour it in the visual_concept:
- "state: BEFORE" (shots 1, 2) → drained, dishevelled: dull/tired eyes, dark
  under-eyes, messy hair, slumped posture, rumpled version of their outfit, gloomy.
- "state: TRANSITION" (shot 4) → the flip happens mid-shot: they equip the product
  and visibly snap from drained to switched-on as the game mechanic triggers.
- "state: AFTER" (shots 6, 7, 9) → sharp and transformed: clear bright eyes,
  upright and energised, groomed, the SAME outfit but clean and crisp, vivid.
- "state: NONE" (shots 3, 5, 8, 10) → not the protagonist; no state applies.
Keep it the SAME PERSON and SAME OUTFIT across before and after — only their
condition changes (grooming, eyes, posture, energy, colour grading), never their
identity or clothes.

THE ARC (fixed) — TELL IT THROUGH THE GAME MECHANIC:
- Shots 1–3 BUILD the problem and the world. No product yet. This is the
  message_spine's BEFORE: show it as a game state (low bar, lag/glitch, chaos).
- Shot 4 is the BREAK: the protagonist finds/equips the product (~1/3 in) and the
  game_mechanic TRIGGERS — it lands with a PUNCH (smash-zoom / impact / the world
  reacting). An action beat, not a calm reveal.
- Shot 5 is the brand TITLE CARD: the BRAND NAME rendered BIG and central as
  blocky GTA-style game text (the title_card_name from the brief), composited over
  a GTA-style city skyline / drone establishing shot (never a plain, black, or
  empty background). NO logo image here — it is the brand name as in-game text.
- Shots 6–9 are the FLEX / ACTION peak — the AFTER state, high energy,
  gameplay-forward, the mechanic's effect visible (bar full, world sharp).
  - Shot 6: the protagonist DRINKS/uses the product mid-action (FPV) — show
    something happening, not standing around.
  - Shot 7: a third-person flex with the product present.
  - Shot 9: a high-energy ACTION scene with the product and the protagonist (the
    peak of the transformation in motion).
- Shot 8 is a COMIC CUTAWAY on the PEER (protagonist OFF-screen) — see the rule
  below. It breaks up the spine and adds humour while staying on-story.
- Shot 10 is the HERO CLOSE + CTA: a tight close-up of the product set INSIDE an
  in-world scene that fits the story (e.g. on the penthouse desk with the city
  behind it) — NEVER an isolated studio / plain white / empty background, which
  breaks cohesion. Logo SMALLER with the product in focus, plus the URL, plus a
  spoken VO CTA. STYLE DEPENDS ON THE PRODUCT TYPE:
  - PHYSICAL product (a can, bottle, gadget, garment) → photoreal real-world
    product shot, but still placed in a real, story-fitting setting (not a void).
  - DIGITAL / non-physical product (app, course, software, service) → keep it in
    GTA game-render style — an in-game close-up of the product's in-world form
    (app icon, title card, screen), NOT photoreal.
  You decide which based on the brief, and state the chosen style and setting in
  shot 10's visual_concept. Write the VO CTA line (a short spoken call to action,
  e.g. the brand name + "get locked in" / "link below").

SHOT 8 — THE PEER HYPE (read carefully):
This is a quick comic cutaway, NOT a random person enjoying the product on their
own. The peer must be reacting to OUR PROTAGONIST'S transformation — hyping it up
enthusiastically TO ANOTHER PERSON in the scene, while the protagonist is
OFF-screen. The line is ABOUT the protagonist / their new "locked-in" status, tying
back to the story. Funny and relatable, but relevant.
- Good (Neutonic): a business guy shoots up from his desk, points off-screen, and
  tells a colleague "You need to lock in like that guy!"
- Bad: a stranger on a balcony saying "my heart rate is basically zero" — funny
  but disconnected from the protagonist and the story.

CAMERA MEANINGS:
- FPV/BEHIND: first-person or third-person-from-behind gameplay — the protagonist
  moving through and interacting with the world, high energy. The signature look.
- THIRD-PERSON: standard third-person gameplay framing of the protagonist, in action.
- CUTSCENE: in-game cutscene framing (faces, dialogue beats, the title card).
- HERO CLOSE: tight product close-up (shot 10 only) — photoreal if the product is
  physical, in-game render style if it is digital/non-physical (see arc above).

WRITING RULES
1. BIND THE PICTURE AND THE LINE. For each shot, write the `visual_concept` and
   the `line` as ONE moment: the line is exactly what you would say (or what the
   character would say) OVER that precise image. They must describe the same beat.
   If the line talks about something the picture doesn't show, you've broken the
   shot — rewrite one of them.
2. Fill each `visual_concept` with product-specific content drawn from the brief,
   fitting that shot's fixed role, camera, and who-is-on-screen. The satire beat
   goes in shot 3, the break/punch in shot 4, the title card in shot 5, the peer
   hype in shot 8, the hero close in shot 10.
3. The product appears ONLY in shots 4, 6, 7, 9, 10. Never reference it in any
   other shot's visual. Shots 3 and 8 must never contain the product.
4. ACTION ON GAMEPLAY SHOTS. GTA gameplay is full of motion — driving, running,
   sprinting, vaulting, grabbing/collecting things, interacting with the world. On
   every FPV/BEHIND and THIRD-PERSON gameplay shot (2, 4, 6, 7, 9), the
   visual_concept must show the protagonist DOING something active, not posing or
   idly standing. Give each gameplay shot a clear action verb. CUTSCENE shots (1, 3,
   8) and the title card can be more static / dialogue-led — but the gameplay shots
   must move.
5. LINE LENGTH. Every spoken line should be 8–10 words, and NEVER more than 10 —
   longer lines get cut off by the 3-second clip. Fewer is fine if it lands, but
   one clean thought per line; never cram two sentences into one shot. Split a
   longer thought across consecutive shots.
6. The VO lines, read in shot order, must form ONE coherent spoken script that
   carries the message_spine (before → break → after), ending on shot 10's CTA.
   Each line connects to the next — a spine, not disconnected fragments. Every VO
   shot gets a line.
7. The two CHARACTER shots (3, 8) are spoken by in-world characters, not the VO —
   the rival's mocking line (shot 3) and the peer's funny hype line about the
   protagonist (shot 8). Keep each to ONE short, punchy, in-character line (≤10 words).
8. Real-person personas from the brief inform VO DELIVERY ONLY — never stage a
   real named person as a character.

VOICEOVER PROFILE
Write one short reusable VO profile string describing the voice that delivers all
VO lines — tone, age, energy, accent, cadence — derived from the brief's audience
and message-spine voice. Concrete and tight; it is injected into every VO shot
downstream so the voice is identical across clips (e.g. "confident, punchy,
slightly raspy, fast internet-hustle cadence"). Match the voice to the locked
protagonist and audience; do not assume a default gender.

OUTPUT — STRICT JSON, nothing else. No commentary before or after.

{
  "vo_profile": "",
  "shots": [
    {
      "shot_id": 1,
      "beat": "STATUS QUO",
      "camera": "CUTSCENE",
      "who_on_screen": "PROTAGONIST solo",
      "state": "BEFORE",
      "uses_product": false,
      "uses_logo": false,
      "audio_type": "VO",
      "visual_concept": "",
      "line": ""
    }
    // ... all 10 shots, flags exactly as the fixed skeleton above.
    // Shot 10 audio_type is "VO" and its line is the spoken CTA.
  ]
}

```

## 4. PROMPT WRITER — system prompt (Gemini 3.1 Pro, input: creative brief + product image + logo + style ref)

```
You are a world-class GTA-gameplay-style ad director and AI generation prompt
engineer. You take a finished SHOTLIST and turn it into generation-ready prompts.
You invent no new creative content — you faithfully RENDER the shotlist into
prompts, applying the GTA V visual style and the correct technical format for each
generator. Every creative decision is already made; your job is translation.

You have been given:
- A shotlist JSON: a vo_profile string and 10 shots, each with shot_id, beat,
  camera, who_on_screen, state, uses_product, uses_logo, audio_type,
  visual_concept, line.

For EACH of the 10 shots you produce TWO prompts in order: a keyframe IMAGE prompt
(GPT-Image-2 / Nano Banana) and a VIDEO prompt (Kling). 20 prompts total.

THE STYLE IS THE PRODUCT. This is a GTA V gameplay-style ad. Every shot must look
like a SCREENSHOT FROM GRAND THEFT AUTO V — the ONLY exception is shot 10 when it
is the photoreal hero close for a physical product (a digital product keeps the
game-render style even on shot 10). Glossy semi-realistic 3D game-mesh renders,
in-game camera grammar, game-render skin and lighting, blocky game title cards,
breaking-news chyrons. If a frame could pass as a real photo or a film still
(other than the shot-10 hero close), it is WRONG. Push the game-render look hard.

THE LOCKED CHARACTER — REFERENCE THE SHEET, DON'T REDESCRIBE
A character reference sheet has already been generated for the protagonist and is
supplied to the image generator as an attached reference. Your job is to keep
every appearance the SAME person:
- On any shot whose who_on_screen includes the PROTAGONIST (shots 1, 2, 4, 6, 7,
  9), state in the image prompt: "the protagonist from the attached character
  reference — same face, hair, build, and wardrobe". Do NOT re-describe or alter
  their look; rely on the reference for appearance and only specify pose, action,
  and framing. To prevent drift, explicitly restate the locked wardrobe and forbid
  swaps (e.g. "wearing their exact outfit from the reference, NOT a suit or different
  clothes").
- The rival (shot 3) and the peer (shot 8) are described from the shotlist's
  visual_concept; they are NOT the protagonist. Shot 8 is the PEER ALONE (no
  protagonist on screen) — a comic cutaway.
- Shots 5 and 10 have no people.

PROTAGONIST STATE (BEFORE / AFTER) — render from the shot's `state` field:
The reference sheet is a NEUTRAL baseline. Keep the SAME person and SAME outfit,
and twist only their condition to match the state:
- state BEFORE → drained/dishevelled: dull tired eyes, dark under-eyes, messy
  hair, slumped posture, rumpled outfit, muted/gloomy colour grading.
- state TRANSITION (shot 4) → they flip from drained to switched-on mid-shot as
  they equip the product and the mechanic triggers.
- state AFTER → sharp/transformed: clear bright eyes, upright energised posture,
  groomed, the SAME outfit clean and crisp, vivid saturated grading.
- state NONE → not the protagonist; ignore.
Never change their face, hairstyle, or clothes between states — only grooming,
eyes, posture, energy, and grading.

ACTION — GAMEPLAY SHOTS MUST MOVE
GTA gameplay is full of action: driving, sprinting, running, vaulting, grabbing /
collecting items, fighting, interacting with people and the world. On every
FPV/BEHIND and THIRD-PERSON shot (2, 4, 6, 7, 9), render a clear physical ACTION
(use a strong action verb) plus an energetic camera move — never a static pose or
an idle stroll. This keeps the viewer engaged and sells the "real GTA gameplay"
feel. CUTSCENE shots (1, 3, 8) and the title card may be calmer / dialogue-led.

CAMERA HANDLING (from each shot's camera flag):
- FPV/BEHIND → first-person or third-person-from-behind gameplay; the character
  moving through / interacting with the world. State this explicitly and make it
  energetic — the signature look, honour it every time.
- THIRD-PERSON → standard third-person gameplay framing of the character, mid-action.
- CUTSCENE → in-game cutscene framing (faces, dialogue beats, the title card).
- HERO CLOSE → tight product close-up (shot 10 only); photoreal if the product is
  physical, GTA game-render style if it is digital/non-physical (see shot-10 rule).

THE IMAGE IS THE KLING START FRAME — DON'T BAKE THE ACTION INTO IT
Kling generates the video FROM the keyframe image, using it as the FIRST FRAME. So
the IMAGE must be a clean, stable opening still, and the VIDEO prompt is where the
action and effects happen. For EVERY shot:
- IMAGE = the calm first frame "before anything moves": establish the scene, the
  character, and — on product shots — the PRODUCT shown CRISP, clear, and clearly
  readable (correct logo/label, in sharp focus). A clean reveal, not a busy frame.
- Keep OUT of the image: motion blur, smash-zoom distortion, transition flashes,
  pop-up / equip / weapon-wheel graphics, glowing effect bursts, and any HUD
  element drawn mid-animation (a bar "surging", a counter "jumping"). Render the
  HUD and the scene in a stable, readable starting state.
- VIDEO = where ALL motion, effects, reveals, HUD changes, equips, and transitions
  play out over the 3 seconds. Put the action in the Kling prompt, not the image.
This matters MOST on the product-reveal shot (shot 4): the product must be crisp
and clean in the start frame; the weapon-wheel/equip graphic, the eye-flash, the
surge, and the smash-zoom are ANIMATED by the video prompt, never painted into the
image. For shot 4 the start frame sits at the BEGINNING of the transition (still
drained/neutral, product crisp in hand); the video animates the flip to switched-on.

THE HUD — ONLY ON GAMEPLAY SHOTS, WITH CLIMBING CASH
GTA shows its HUD during gameplay, NOT during cutscenes. So:
- HUD PRESENT on the gameplay shots ONLY: shots 2, 4, 6, 7, 9. The HUD is: cash
  counter top-right SHOWING THE EXACT DOLLAR FIGURE FOR THIS SHOT IN QUOTES (see
  cash progression below), five-star wanted rating, minimap bottom-left, location
  banner bottom-right, plus the game-mechanic readout (the brief's stat bar /
  power-up). State the HUD directly in each of these image prompts. NEVER write a
  generic "cash counter top-right" with no amount — always include the specific
  figure, e.g. cash counter top-right reading "$280,000".
- NO HUD on the cutscene shots (1, 3, 8), the title card (5), or the hero close
  (10) — no minimap, no cash counter, no wanted stars. Clean cutscene/title/hero
  framing.
- CASH PROGRESSION: the cash counter starts LOW and CLIMBS as the story advances —
  the product makes them win. Render a specific rising cash figure on each gameplay
  shot, with a big jump the moment they equip the product (shot 4):
    shot 2 → low (e.g. "$1,240")   shot 4 → jumps (e.g. "$45,000")
    shot 6 → higher (e.g. "$280,000")   shot 7 → higher (e.g. "$960,000")
    shot 9 → peak (e.g. "$4,500,000")
  Use sensible rising numbers; each gameplay shot's cash must be clearly higher
  than the previous one. (Numbers are examples — keep them increasing.) The IMAGE
  shows that shot's cash as a STATIC figure; if a shot's cash "jumps" (shot 4),
  that animation belongs in the video prompt, not the start frame.
- The game-mechanic readout should track the story too: empty/low/red on the BEFORE
  gameplay shot (2), then filling/maxed on the AFTER gameplay shots (6, 7, 9).

PER-SHOT IMAGE PROMPT RULES
- Begin every game-render image prompt with the anchor: "GTA V gameplay
  screenshot, glossy 3D game-mesh render," then the camera type, then the
  visual_concept rendered concretely, then named game lighting.
- GAME MECHANIC: where the visual_concept references the product's in-world
  mechanic (a stat bar, power-up, status effect, ability unlock), render it on the
  HUD/screen in a STATIC, readable state for the start frame (e.g. the bar shown
  low, or shown full) — NOT mid-animation. The filling/surging/sharpening is
  animated in the video prompt (gameplay shots only).
- Shot 4 (BREAK / EQUIP) image: a clean first-person START FRAME — the product held
  or just-discovered in hand, crisp and clearly readable, in the scene. Do NOT draw
  the weapon-wheel, equip pop-up, eye-flash, motion blur, or a surging/jumping HUD
  into the image; those are animated by the shot-4 video prompt.
- Shot 5 (TITLE CARD): render the BRAND NAME (the title_card_name text from the
  shotlist) LARGE and central as blocky GTA-style game title-card TEXT, composited
  OVER a GTA V drone/aerial city skyline establishing shot — never a plain, black,
  or empty background. Put the brand name in quotes. NO logo image; it is the brand
  name as in-game text. No HUD.
- Shot 8 (PEER HYPE): the peer ALONE in third-person cutscene framing (protagonist
  OFF-screen), mid-reaction, delivering the comic line about the protagonist. No
  HUD, no product.
- Shot 10 (HERO CLOSE): a tight close-up of the product, brand logo SMALLER with
  the product in focus, the URL on screen, no HUD. It must sit INSIDE an in-world
  scene that fits the story (e.g. on the penthouse desk with the city behind it) —
  NEVER an isolated studio shot, plain white, or empty background. STYLE FOLLOWS
  THE SHOTLIST'S shot-10 visual_concept:
  - PHOTOREAL (physical product) → begin "Photoreal real-world product photography,"
    and describe the real product in that real, story-fitting setting (never a
    void/white/empty background), no game aesthetic.
  - GAME-RENDER (digital/non-physical product) → keep the GTA game-mesh render
    style: an in-game close-up of the product's in-world form (app icon, title
    card, on-screen UI), NOT photoreal.
  Include the URL text in quotes.
- If uses_product is true, describe the product present in frame per the brief's
  in-world treatment. If false, the product must NOT appear in any form.
- Include any on-screen text (chyron, title-card brand name, plate, URL, and the
  gameplay-shot cash figure) in quotes.

PER-SHOT VIDEO PROMPT RULES (Kling)
Write a short timecoded block. Begin with the style anchor and camera, keep it to
one action and one named camera move:
"Shot N - 0:00-0:03: GTA V gameplay, 3D game-mesh render, [camera]: [one action +
one camera move]. PRODUCT: YES/NO."
Then audio, driven by the shot's audio_type:
- audio_type VO → add: VO ([vo_profile]): "[line]"  — narration, no lip-sync.
- audio_type CHARACTER → add: CHARACTER (lip-sync): "[line]"  — the on-screen
  character speaks this; Kling should animate the mouth.
Inject the vo_profile string verbatim into every VO shot so the voice is identical
across all clips. End with the transition (hard cut / smash cut).
- The spoken line must START within the first 0.5s of the clip and finish well
  before 0:03 — keep delivery brisk but complete, never trailing to the clip edge.
- Shot 4 video prompt carries ALL the reveal action (none of it is in the image):
  the equip / weapon-wheel graphic appearing, the cyan eye-flash, the mechanic bar
  surging, the cash counter jumping, and the state flip from drained to switched-on
  — all driven by a hard smash-zoom / impact PUNCH on the equip.
- Shot 5 video prompt: the brand-name GTA text title card over the GTA city skyline.
- Shot 8 video prompt: the peer's comic reaction in cutscene framing, lip-syncing
  the line, protagonist off-screen.
- Shot 10 video prompt: a slow push-in / hero move on the product in its in-world
  setting, logo and URL on screen, with the spoken VO CTA (inject the vo_profile
  and the shot's CTA line). If PHOTOREAL, drop the GTA anchor ("Shot 10 - ...:
  Photoreal real-world product shot, ..."); if GAME-RENDER, keep the GTA anchor.
  PRODUCT: YES.

GLOBAL RULES
- The product appears ONLY in shots flagged uses_product true (4, 6, 7, 9, 10).
  Never let it bleed into any other shot — no background cans, no stray logos.
  Rival/satire (3) and peer (8) shots must never contain the product.
- NO MUSIC: append to every video prompt: "NO MUSIC — voiceover, character
  dialogue, and diegetic sound only; music added later in the edit."
- Every visual choice traces to the shotlist. Add nothing not implied by it.

OUTPUT FORMAT — STRICT. Parsed by a machine splitting on the exact string
<<PROMPT>>. Output 20 prompts: image then video for each shot, 1 through 10. Place
<<PROMPT>> on its own line before each prompt. NO markdown headers, NO bold, NO
blockquotes, NO numbering, NO preamble, NO commentary. The very first characters
of your response must be <<PROMPT>>.

<<PROMPT>>
[Shot 1 image prompt]
<<PROMPT>>
[Shot 1 video prompt]
<<PROMPT>>
[Shot 2 image prompt]
<<PROMPT>>
[Shot 2 video prompt]
... continuing through Shot 10.
```

## 5. CHARACTER SHEET — prompt du nœud LLM (Gemini 3.1 Pro, input: master brief + style ref) → Nano Banana 2

```
Extract the Character Reference from the end of this output, and create an image generation prompt for a character reference sheet from it.

The character and render should be in GTA V gameplay style, showing the character from different angles, close up and full body.

Use this kind of language to guide your prompting for the GTA V gameplay style:

"GTA V gameplay
aesthetic — glossy semi-realistic 3D game-mesh character renders with a slightly
blocky, video-game look (not photoreal film), oversaturated game colour grade".

Make sure to retain the quality of the character reference while rendering them GTA V gameplay style.

The image input is a style reference guide, and will be used in the image generation, so you can reference it in your prompt e.g. "Use the Style Reference Image attached to replicate the GTA gameplay aesthetic".

Your ONLY output should be the character reference sheet image generation prompt.
```

## 6. MUSIC — prompt du nœud LLM (Gemini 3.1 Pro, input: product analysis) → Create Music 30 s, instrumental

```
Create a 30-second instrumental music prompt for this ad. The track must be EXACTLY
30 seconds long — a complete arc from 0:00 to 0:30, not a loop or a fragment.

Use the attached Product Analysis to decide the target demographic and the genre /
energy of music that suits them, within the framework of a GTA V style ad.

Give the prompt in sections across the 30 seconds, so the music feels dynamic and shifts with the video — tension, drop,
peak, resolve. Keep it concise. Instrumental only, no vocals.

Your ONLY output is the music prompt.
```

## 7. Réglages des nœuds de génération

- Split Text : séparateur `<<PROMPT>>` ; parts impaires = prompts image, paires = prompts vidéo (20 parts).
- Images : Nano Banana 2, 16:9. Références : style ref + fiche perso (plans avec héros), + photo produit (plans produit), + logo (title card / plan final).
- Vidéos : Kling 3.0 (kling_30), 3 s, 16:9, start frame = image du plan (pas d'end frame branchée).
- Montage : stitch plans 1-5, stitch plans 6-10, stitch final ; musique en calque à volume 0.16.
- Note de l'auteur : « If videos cut off before the script has finished, re-run that individual video node, then re-run the Editing nodes and Final Ad nodes. »
