UGC

Two-stage workflow:

1. UGC image — ultra-realistic iPhone front-camera selfie at 9:16 2K
2. UGC video — talking-head video from the selfie at 9:16 1080p via Veo 3.1 Fast

Output saved to ./brands/[brand-name]/ugc/[output-name]/.

Step 1 — Select brand

Scan ./brands/ for subfolders that contain a brand-identity/visual-guidelines.md file.

• One brand found: use it automatically and confirm: "Using brand: [name]"
• Multiple brands found: list them and ask which to use
• None found: tell the user to run /brand first

Step 2 — Select a model

Ask: "Would you like to use a styled model (specific outfit applied) or an unstyled model (plain casting neutral wardrobe)?"

If styled: list available outputs from ./brands/[brand-name]/styled/:

Available styled models:
  • jade-cloudmerino-tshirt
  • marcus-techsilk-shorts

Read brands/[brand-name]/styled/[name]/style-spec.json to get model_dir.

If unstyled: list available models from ./brands/[brand-name]/models/:

Available models:
  • jade
  • marcus-02

If neither folder has completed outputs:

"No models found for [brand]. Use /model to create one first. If you want a specific outfit, also run /style."

Step 3 — Collect variables

Ask for all in one message:

Variable 1 — Action: "What is the model doing? (e.g. standing, sitting at a café, walking down a street, leaning against a wall)"

Variable 2 — Location: "Where are they? Describe the environment in detail. (e.g. a busy urban high street with bokeh traffic in the background, a sunlit café terrace in Paris, a graffiti-covered alley at golden hour)"

Variable 3 — Outfit (unstyled models only): "What are they wearing? Describe the full outfit. (e.g. a black oversized hoodie, dark jeans, white sneakers)"

Skip Variable 3 for styled models — outfit is locked to the styled image.

Step 4 — Name the output

Derive a slug from the model name and a short descriptor:

• jade-paris-cafe
• marcus-urban-walk
• hero-golden-hour

Ask the user to confirm or rename.

Step 5 — Write ugc-spec.json

Create the output folder and write the spec:

brands/[brand-name]/ugc/[output-name]/
  ugc-spec.json

ugc-spec.json — styled model:

{
  "output_name": "jade-paris-cafe",
  "brand": "[brand-name]",
  "model_type": "styled",
  "model_dir": "brands/[brand-name]/models/jade",
  "styled_dir": "brands/[brand-name]/styled/jade-cloudmerino-tshirt",
  "action": "sitting at an outdoor café table, leaning slightly forward",
  "location": "A sunlit café terrace in Paris, warm afternoon light, blurred street and pedestrians behind",
  "outfit": "must perfectly match image 2"
}

ugc-spec.json — unstyled model:

{
  "output_name": "marcus-urban-walk",
  "brand": "[brand-name]",
  "model_type": "unstyled",
  "model_dir": "brands/[brand-name]/models/marcus",
  "styled_dir": null,
  "action": "walking down a street, mid-stride",
  "location": "Busy urban high street, motion-blurred cars and shop fronts in background",
  "outfit": "A black oversized hoodie, dark jeans, white sneakers"
}

Step 6 — Generate image

Run from the project root:

python3 skills/references/generate-ugc.py brands/[brand-name]/ugc/[output-name] --image

Reference images passed in order:

• Image 1: brands/[brand-name]/models/[name]/headshot.png — identity
• Image 2 (styled only): brands/[brand-name]/styled/[name]/[name]_v{n}.png — outfit (script auto-detects the latest version)

Cost: ~$0.12

Step 7 — Review and decide

Tell the user the exact path where the file was saved, e.g.: brands/[brand-name]/ugc/[output-name]/[output-name]_v1.png

Ask: "Happy with this image? You can:

• Continue — pick a version to use for the video, then move to Step 8
• Regenerate — run again to get another version
• Adjust — change the action, location, or outfit and regenerate"

If more than one version exists when the user is ready to continue, ask which they want to use for the video.

To regenerate:

python3 skills/references/generate-ugc.py brands/[brand-name]/ugc/[output-name] --image

Step 8 — Video: collect inputs

Once the user is happy with the image, ask:

Script: "What should they say in the video? Paste the script."

Voice and delivery notes: "Any notes on accent, tone, pace, or energy? (e.g. 'Australian accent, upbeat and conversational', 'calm and authoritative, slight American accent')"

Duration: "How long? Options: 4s, 6s, or 8s (Veo 3.1 Fast only supports these three values)"

Realism patterns (apply when composing the video prompt)

Before writing the final Veo prompt, read skills/references/shlabu-realism.md and apply its principles to the way you frame the script + delivery prompt. The Veo-specific suffix in that reference is paste-ready and should be appended to the delivery notes for any selfie/UGC clip aiming for "raw iPhone" realism.

Key levers for UGC video (full reference for context):

• Repeat anti-synthetic vocabulary ("handheld, organic shake, unhurried, not performing, just messing around on her phone") 4–6× across the prompt
• Call out high-frequency skin/hair detail ("visible pores, scattered freckles, baby hairs, individual strands catching the light")
• Describe the background as deliberately soft / blurred so Veo doesn't waste detail there
• Quote each spoken line with delivery direction inline — "line" — tone, gesture, beat

Step 9 — Update ugc-spec.json with video inputs

Add video fields to the existing spec:

{
  ...existing fields...,
  "script": "I've been obsessed with this tee all summer. It's the CloudMerino 66 from Satisfy — lightweight, breathable, and it just feels incredible.",
  "voice_notes": "Australian accent, relaxed and genuine, like talking to a friend. Not salesy.",
  "duration": "8"
}

Step 10 — Generate video

Run from the project root:

python3 skills/references/generate-ugc.py brands/[brand-name]/ugc/[output-name] --video

Uploads ugc-image.png to FAL, combines script + voice notes as the prompt, calls Veo 3.1 Fast at 9:16 1080p.

Cost: ~$0.60 for 4s · ~$0.90 for 6s · ~$1.20 for 8s at 1080p with audio.

Step 11 — Present the result

Tell the user the exact path where the video was saved, e.g.: brands/[brand-name]/ugc/[output-name]/[output-name]-video.mp4

Ask: "Happy with the video? Or would you like to regenerate, adjust the script, or change the delivery notes?"

To regenerate video:

python3 skills/references/generate-ugc.py brands/[brand-name]/ugc/[output-name] --video

To run image + video together:

python3 skills/references/generate-ugc.py brands/[brand-name]/ugc/[output-name] --image --video

Notes

• Each brand has its own ugc/ subfolder — multiple brands can run independent UGC workflows in the same project
• UGC image is always 9:16 to match the video format
• generate_audio: true is set — Veo 3.1 Fast generates audio alongside the video
• Duration is locked to 4, 6, or 8 seconds — Veo 3.1 Fast does not support other values
• The face in the video is driven by the UGC image. A strong, well-lit selfie = more consistent video output
• If the video doesn't match the script delivery, regenerate — Veo output varies by seed