Creating AI UGC Video Ads

Given a product and audience, produce a 15-45s vertical UGC-style ad:

1. Lock brief: one goal, one persona, one platform, one length.
2. Write script: Hook (0-3s) → Body (problem/discovery/result) → CTA. Spoken language only.
3. Shot-list it: one action per shot, mix selfie talking-head + product close-up + B-roll + reaction.
4. Generate or film: style bible locked (face/wardrobe/setting), first-frame approved before animating, real product composited in post if label text garbles.
5. Edit: jump cuts every 2-4s, burned-in casual captions, light grain, no cinematic polish.
6. QA: mute test, phone test, stranger test. Regenerate near-misses.
7. Compliance: disclose AI/sponsored content, verify claims, confirm likeness rights.

Default framework when no brief exists: Problem → Discovery → Result, because it reads as the most authentic testimonial arc.

Progress:

• Step 1: Lock the brief (goal, persona, platform, length, constraints)
• Step 2: Write script (hooks, body, CTA — spoken language, timed at 2-3 words/sec)
• Step 3: Build shot list (one action per shot, varied shot types)
• Step 4: Capture or generate footage (filmed or AI pipeline)
• Step 5: Audio pass (real or synthetic voice + room tone + minimal/no music)
• Step 6: Edit and finish (cuts, captions, grading, safe zones)
• Step 7: QA (artifact check, mute test, phone test, stranger test)
• Step 8: Compliance check (disclosures, claims, likeness)
• Deliver final package

Step 1 — Brief

Collect: product/packaging/claims, audience pain points (mine Amazon reviews, Reddit, competitor ads), creator persona (age/vibe/setting/plausibility), single objective (sell/demo/review/unbox/testimonial), platform + aspect ratio (9:16) + length (15-45s).

Step 2 — Script

• Framework options: Hook-Body-CTA (default), PAS (pain-point products), AIDA (longer demos).
• Write 3-5 hooks, 2-3 body angles, 1-2 CTAs → combine for cheap variant testing.
• Spoken language: contractions, fragments, filler words, imperfect grammar.
• Specific details beat adjectives ("my skin stopped peeling in a week" > "amazing results").
• Claims must be backable and FTC-compliant.
• Pace: ~2-3 spoken words/second.

Step 3 — Shot list

One action per shot. Shot-type mix: selfie talking-head, product close-up, use B-roll, reaction shot. Plan jump cuts every 2-4 seconds.

Step 4 — Capture or generate

Filmed: handheld phone framing, natural/imperfect lighting, lived-in environment, authentic product handling, clean voice + room tone recorded separately.

AI generation:

• Lock a style bible (face, wardrobe, setting descriptors) — paste verbatim into every prompt.
• Prompt formula: subject + action + setting + camera + lighting + mood.
• Add realism descriptors: "shot on iPhone, handheld, natural window light, slight grain, unretouched skin, cluttered background."
• Generate/approve first-frame still → animate via image-to-video, one action per clip, simple camera moves only.
• Chain continuity using last-frame-of-previous-clip as first-frame-of-next.
• Use negative constraints where supported (no studio lighting, no perfect skin, no text overlays).
• Composite real product/logo in post if the model distorts label text.
• Model choice depends on need: Seedance/Veo/Kling/Sora/Runway/Luma/Pika for video; Midjourney/Flux/Ideogram/Stable Diffusion/GPT-image for keyframes; HeyGen/Synthesia/Hedra/Captions for talking-head lip-sync when native video-model sync is weak.

Step 5 — Audio

Prefer real voice. If synthetic (ElevenLabs), write stumbles/breaths into the script. Light noise reduction only (Adobe Podcast Enhance) — overprocessed = fake. Layer faint ambient room tone. Music low in mix or omitted.

Step 6 — Edit and finish

CapCut (native feel) / Premiere / DaVinci / Descript. Jump cuts every 2-4s. Casual burned-in captions, word-level highlight style. Subtle grain, light handheld shake/reframe, light compression, no cinematic LUTs. Cover weak shots with B-roll cutaways. Keep text/faces inside platform safe zones.

Step 7 — QA

Check: hands/fingers, teeth, eye gaze, label/text legibility, lip-sync drift, morphing objects between frames. Run mute test, real-phone test, stranger test ("would this pass as a real customer?"). Regenerate near-misses (80%-right shots) rather than patching in post.

Step 8 — Compliance

Apply platform AI-generated and sponsored-content labels. Verify claims against FTC endorsement guidelines and ad-platform policy (especially health/finance/before-after claims). Confirm no real person's likeness/voice used without consent.

Example 1: Input: Skincare serum, target audience is women 25-40 with eczema-prone skin, objective = testimonial ad for TikTok. Output: Problem→Discovery→Result script ("I was so sick of my skin flaking every winter..." → "a friend told me to try this..." → "a week in and it's calmed down completely"), persona = 32-year-old woman filming in her bathroom at night with lamp light, shot list of selfie talking-head + product close-up + hand applying serum + after-shot reaction, AI-generated using locked style bible if not filmed, burned-in captions, 28-second runtime, hook: "My dermatologist did NOT tell me about this."

Example 2: Input: AI-generated unboxing video for a tech gadget, Sora or Kling pipeline, model garbles the logo text. Output: Generate full unboxing scene with product placeholder, then composite real product photo/logo onto the package in post-editing; keep one action per shot (open box → turn product → reaction) to minimize warping; regenerate the "hands opening box" shot twice due to finger artifacts before accepting.