Google Flow UGC Ad Generator
Generates complete production packages for multi-shot UGC video ads in Google Flow. One product image plus one ad angle produces everything the user needs to run a full Agent mode session in Flow — creator prompt, shot list, tagged prompts, Agent instruction, and voiceover script.
What this skill does
- Takes a product image path and ad angle from the user
- Optionally reads
brand/brand-dna.md,brand/brand-voice.md,brand/icp-cards.mdif they exist — skips silently if not - Generates a realistic AI creator image prompt optimized for ChatGPT Images 2.0 or Nano Banana Pro
- Writes a 5-shot direct response script with dialogue for each scene
- Produces five asset-tagged Flow prompts ready to paste directly into Flow
- Writes a single Agent mode briefing instruction covering the full production session
- Produces an ElevenLabs-ready voiceover script for post-production audio
This skill does NOT call any APIs. It produces the production package. The user runs the actual generation in Flow's browser UI.
Before you start
The user needs:
- A Google AI Pro subscription ($19.99/month) at labs.google.com/flow
- A product image saved locally (JPG or PNG)
- Optionally: an existing creator image, or they generate one using the character prompt this skill outputs
No API keys. No code. No terminal. This is a UI tool.
Workflow
Step 1 — Intake
Collect from the user:
- Product image — file path or description (required)
- Product name + one-line description — if not provided, ask
- Ad angle — pick one or ask:
testimonial— creator speaks to camera about results (default)car/on-the-go— creator in car or commuting, convenience angleunboxing— creator reveals and reacts to productlifestyle-demo— creator using product in natural environmentproblem-solution— before state, product intro, after state
- Creator description — brief description (age, gender, vibe). If not provided, default to: "27-year-old woman, relatable fitness-oriented, not model-perfect"
- Brand/product category — supplement, beauty, apparel, food/bev, home goods, etc. Informs which angle and which creator archetype fits best.
Do not ask for everything at once. Product image + angle is enough to start. Ask for missing details naturally.
Step 2 — Check for brand context
Look for:
brand/brand-dna.mdbrand/brand-voice.mdbrand/icp-cards.md
If any exist, read them silently. Use them to sharpen dialogue voice and ICP targeting. Do not mention this to the user. If none exist, proceed — this skill works with just a product image.
Step 3 — Generate the creator image prompt
Output a character image prompt the user can paste into ChatGPT Images 2.0 or Nano Banana Pro.
Follow this formula exactly:
Candid iPhone photo of a [age]-year-old [gender], [specific physical detail — hair color/texture with natural flyaways, freckles or not, etc.], [minimal/natural/no] makeup, wearing [specific casual outfit — real clothing, not "stylish"], [environment detail — sitting in car / standing in kitchen / etc.]. Shot from [slight angle — low angle slightly off-center / slightly above eye level]. [Specific lighting — soft overcast daylight through window / warm morning kitchen light / overhead fluorescent in gym]. Slightly imperfect exposure. Real, unretouched skin texture. Natural hair flyaways. No studio lighting. Editorial-style realism.
Rules:
- Never use "beautiful," "gorgeous," "stunning," or model-descriptors
- Always include a specific environment, even in the character reference shot
- Always include at least one deliberate imperfection (flyaways, freckles, uneven lighting)
- Always include "Candid iPhone photo" — one of the strongest realism signals
- Match the creator archetype to the product category (see
references/characters.md)
Tell the user: save the generated image as creator.png and upload it to their Flow project along with their product image. Both filenames should be descriptive: creator.png and [product-name].jpg.
Step 4 — Write the 5-shot DR script
For the chosen angle, write a complete shot list. Each shot gets:
Shot N/5 — [Shot name]
Environment: [where the creator is — specific room/setting/context]
Action: [exactly what's happening — one action only per shot]
Product in frame: [yes/no — and if yes, how]
Dialogue: [what the creator says — ~3 words per second, UGC voice]
Direct response structure (default — testimonial angle):
| Shot | Job | Environment | Product |
|---|---|---|---|
| 1 — Hook | Stop the scroll with a single claim | Kitchen / bathroom / neutral | Yes — held toward camera |
| 2 — Problem | Make the viewer feel the pain | Bedroom / desk / car | No |
| 3 — Discovery | Introduce the product naturally | Kitchen / bathroom | Yes — opening/interacting |
| 4 — Transformation | Show the result. Specific, not vague | Gym / outdoors / work | No |
| 5 — CTA | One action, remove friction | Back to Shot 1 environment | Yes — held toward camera |
Dialogue rules:
- UGC voice: "Okay I need to talk about this" not "Introducing our new product"
- ~3 words per second of screen time. A 7-second shot = ~21 words max
- First line is the hook — must work without sound (captions carry it on autoplay)
- Product name appears in Shot 3 or later, never Shot 1
- CTA is punchy: "link is below" / "you literally have nothing to lose" — never "shop now" or "buy today"
- If
brand/brand-voice.mdwas loaded, mirror its sentence patterns
See references/angles.md for shot-by-shot breakdowns of all five preset angles.
Step 5 — Write the asset-tagged Flow prompts
For each shot, produce a prompt ready to paste into Flow. Every prompt tags both the creator image and the product image by filename.
Prompt structure:
Vertical UGC video, [creator.png] [environment + outfit detail], [action]. [product.jpg if in frame — how they're holding/interacting with it]. Speaking directly to lens: "[dialogue]." [Style modifiers]. 9:16 vertical, 8 seconds.
Required style modifiers for every shot:
handheld feel, slight camera shakecandidnatural skin texture[environment-appropriate lighting — soft window light / overcast natural light / warm kitchen light]imperfect framing
Product-forward shots additionally add:
product label clearly visible[product color] bottleor appropriate descriptor — reinforce the reference visually
Example (Shot 1 — Hook, testimonial angle, supplement product):
Vertical UGC video, creator.png standing in a bright kitchen, auburn hair in loose ponytail, light lavender athletic top, holding goli-bottle.jpg up toward camera with both hands. Product label clearly visible, red bottle. Speaking directly to lens: "I've been taking these every morning for 30 days and my energy is actually different." Handheld feel, slight camera shake, candid, natural skin texture, soft window light, imperfect framing. 9:16 vertical, 8 seconds.
Step 6 — Write the Agent mode briefing instruction
Produce a single instruction the user can paste into Flow's Agent mode to brief the entire session at once.
Format:
I'm creating a 5-shot vertical UGC ad for [product name].
My uploaded assets:
- creator.png — the UGC creator
- [product-filename].jpg — the product
Generate these shots in order:
Shot 1: [environment]. creator.png [action + product interaction]. Speaking to camera: "[dialogue]." [lighting], handheld, candid, imperfect framing. 9:16, 8 seconds.
Shot 2: [different environment]. creator.png [action, no product]. Speaking to camera: "[dialogue]." [lighting], handheld, candid. 9:16, 8 seconds.
Shot 3: [environment]. creator.png [action + product interaction]. Speaking to camera: "[dialogue]." [lighting], handheld, candid. 9:16, 8 seconds.
Shot 4: [active environment]. creator.png [action, no product]. Speaking to camera: "[dialogue]." [lighting], handheld, candid. 9:16, 8 seconds.
Shot 5: [clean environment]. creator.png holding [product-filename].jpg toward camera. Speaking to camera: "[dialogue]." [lighting], handheld, candid. 9:16, 8 seconds.
Keep creator.png's appearance consistent across all shots. Generate each shot at 9:16 vertical.
Step 7 — Write the ElevenLabs voiceover script
Produce the complete voiceover script — all five shots in order, with timing notes.
Format:
VOICEOVER SCRIPT — [Product Name] UGC Ad
Total duration: ~40 seconds
[Shot 1 — 0–8s]
"[dialogue]"
[Shot 2 — 8–16s]
"[dialogue]"
[Shot 3 — 16–24s]
"[dialogue]"
[Shot 4 — 24–33s]
"[dialogue]"
[Shot 5 — 33–42s]
"[dialogue]"
---
ElevenLabs settings: Stability 0.5, Similarity 0.75, Style 0.3
Recommended voice style: conversational, warm, not announcer
Tell the user: generate the video clips in Flow first. Then record or generate this voiceover in ElevenLabs. Stitch both in CapCut — video layer + audio layer + captions.
Step 8 — Deliver
Present the full production package in this order:
- Creator image prompt — paste into ChatGPT Images 2.0 or Nano Banana Pro
- Shot list with dialogue — approve before generating
- Five asset-tagged Flow prompts — one per shot, paste individually
- Agent mode briefing instruction — paste once to brief the full session
- ElevenLabs voiceover script — use after video generation
Ask the user if they want to adjust dialogue or angle before they generate. Dialogue changes are free. Regenerating clips costs credits.
File structure this skill produces
[project-name]-production-package/
├── character-prompt.md # ChatGPT Images / Nano Banana prompt
├── shot-list.md # 5-shot script with dialogue
├── flow-prompts.md # 5 asset-tagged prompts for Flow
├── agent-instruction.md # Single Agent mode briefing
└── voiceover-script.md # ElevenLabs-ready script with timing
References
references/prompting.md— style modifiers, asset tagging syntax, Agent mode briefing structure, what breaks character consistencyreferences/angles.md— shot-by-shot breakdowns for all 5 preset anglesreferences/characters.md— character image prompt templates for 8 creator archetypes
Honest limits
- This skill produces prompts, not video. The user generates video in Flow's browser UI.
- Voice consistency across shots is not possible in Flow without the Avatar feature (own face/voice only). Plan for ElevenLabs post-production.
- Product label text can garble when the product is small in frame or the environment is complex. Keep product large and close in product-forward shots.
- Character consistency in Flow requires tagging
creator.pngin every shot prompt. Missing one tag = creator drift. - Flow's Agent mode is experimental. If it doesn't run all shots automatically, paste individual prompts one at a time.
- Eating/drinking shots hallucinate. Use hold-and-show or hold-and-smile endings for any consumable product.