Ugc Ads

Pica Adapter

This skill has been adapted for Pica. Treat any upstream-specific command snippets as workflow examples, not the current command contract.

• Use pica status before paid work to confirm auth, credits, active environment, and current project.
• Use pica --schema, pica --schema=.generate, pica --schema=.model, and pica --schema=.assets for exact command shapes.
• Never guess model IDs. Run pica model search <query> and pica model info <model-id> before constructing generation input.
• Route media generation through pica generate --input @file.json5; route uploads, downloads, and public URLs through pica assets.
• Keep templates, prompt cookbooks, reference assets, scripts, and examples inside the skill folder as skill-owned assets.
• Do not treat visual review heuristics as hard infrastructure gates. Use them as low-confidence checks unless the issue is mechanically measurable, such as missing files, wrong MIME type, bad duration, missing audio, or invalid dimensions.

What Is UGC Ads

This skill is for performance creative, not generic "make me a video" generation.

Use it when the output is meant to behave like:

• TikTok or Reels paid ads
• creator-style product demos
• testimonial or problem-solution clips
• talking-head + product B-roll hybrids
• fast creative testing with multiple hooks and variants

The goal is not to produce the prettiest video. The goal is to produce a video that feels native enough to stop the scroll and clear enough to sell.

Strategy

UGC ad work fails when the workflow starts with assets instead of the ad angle.

Do not start with "generate an avatar". Start with:

1. audience
2. pain point
3. hook
4. proof
5. CTA

Only after the ad logic exists should you choose models and generate media.

Workflow

Step 1: Lock the brief

Before any generation, clarify:

• product and offer
• target audience
• primary pain point
• single desired action
• platform target (TikTok, Reels, shorts, etc.)
• whether the ad should feel like real creator UGC, AI-avatar UGC, or product-led montage

If these are vague, stop and tighten them. A fuzzy brief produces expensive sludge.

Step 2: Write the ad angle first

Every draft should have a clear structure:

• Hook: first 1-3 seconds
• Problem: what frustrates the audience
• Solution / proof: demo, result, mechanism, social proof
• CTA: what to do next

Do not write one script and declare victory. Start with multiple hook angles.

Read creative-strategy before scripting.

Step 3: Choose the production route

Pick one route before choosing models:

Route	Best when
Creator-led UGC	Real person on camera, authentic testimonial feel matters most
AI avatar UGC	Need speed, localization, or many variants without creator scheduling
Product-led demo	Product visuals and proof matter more than face time
Hybrid	Talking head for hook + B-roll for proof and CTA

Read:

• model-selection for route-to-model mapping
• talking-head for avatar / lip-sync choices
• b-roll-and-product for product visuals and proof shots

Step 4: Research before generating

Before touching a model:

pica prompt find "ugc ad"
pica prompt find "tiktok hook"
pica skill find "talking avatar"
pica skill find "banana prompting"

Then discover current model options:

pica model search "seedance"
pica model search "avatar"
pica model search "lip sync"
pica model search "seedream"
pica model info <model-id>

Never guess current IDs. Search finds candidates; model info confirms the real contract.

Step 5: Build cheap drafts first

Your first pass is for message validation, not polish.

• shortest viable duration
• lowest acceptable resolution
• one hook angle per draft
• only enough B-roll to prove the concept

Generate multiple variants instead of over-perfecting one.

Step 6: Review like a performance marketer

After each draft, ask:

• Is the hook visible in the first 3 seconds?
• Does the product appear early enough?
• Is the claim specific?
• Does the clip feel native or over-produced?
• Is the CTA obvious?

If the answer is "looks cool but not convincing", the creative angle is wrong. Don't blame the model first.

Read testing-and-iteration.

Step 7: Finalize only after direction is confirmed

Once a route works:

• upgrade the winning variant only
• produce 3-5 meaningful creative variants
• localize or re-record voice if needed
• keep packaging native to the target platform

Do not spend premium generations on unvalidated concepts.

Model Philosophy

No single model wins every part of UGC ads.

Use a stack, not a religion:

• image models for faces, products, and reference frames
• kling-avatar or performance models for direct address
• video generation models for B-roll, transitions, and product dramatization

Right now, Seedance 2.0 is a strong direction for ad-style multi-asset video work, but it is not a blanket replacement for talking-head systems or creator footage. Read model-selection before committing.

Production Rules

• UGC ads are usually stronger when they feel native, not cinematic-for-its-own-sake
• Hook clarity beats visual beauty
• Proof beats adjectives
• One ad = one core promise
• Fast testing beats single-masterpiece thinking

When To Read Each Reference

• creative-strategy: ad angle, hooks, script structure
• model-selection: choosing the model stack for each route
• talking-head: avatar, lip sync, presenter performance
• b-roll-and-product: product shots, demos, proof visuals
• testing-and-iteration: review loops, variant strategy
• platform-and-compliance: TikTok / Meta platform fit, commercial UGC constraints
• source-notes: external references and dated notes that informed this v1

Minimal Workflow Example

# 1. Research
pica prompt find "ugc ad skincare"
pica model search "seedance"
pica model search "avatar"
pica model search "seedream"

# 2. Inspect current model contracts
pica model info <video-model-id>
pica model info <avatar-model-id>
pica model info <image-model-id>

# 3. Draft the first cheap variant
pica generate \
  --model <draft-video-model-id> \
  --kind video_generation \
  --input '{
    "prompt": "Vertical UGC-style skincare ad. Hook: \"I thought this was another gimmick until day 5.\" Young woman speaking directly to camera in a bedroom, honest creator tone, subtle hand gestures, product appears within first two seconds, quick cut to close-up product texture demo, clear CTA ending."
  }'

Default Mindset

Think like:

• a direct-response creative strategist when scripting
• a creator when shaping tone
• an editor when structuring pacing
• a media buyer when deciding what to test next

If those four perspectives disagree, the media buyer wins.

Creator-Style UGC Upgrade

The existing Pica workflow is the retained base because it keeps the CLI as infrastructure and lets the skill carry vertical craft. The following production rules were absorbed from the upstream UGC-ad skill and adapted for Pica.

Absorbed UGC-Ad Craft

A generalized overlay, not a finished ad. It tells the pipeline HOW believable UGC ads work so the scenarist and art-director produce a strong one for whatever the user is selling. It runs through the normal pipeline and never bypasses the intake gates, the reference-required gate, or the quality gates.

Hard invariants

• Provider invariants stand. All generation via pica generate (Pica-supported providers only — no FAL, no Higgsfield, no Canva, no raw API). Renders via the user's chosen composition/render path. Read pica model search / pica model info before naming any model id.
• Reference gate. A named real brand product ("iPhone 16", "CeraVe cleanser") fires the reference-required gate — refuse without a ref photo or logged --no-ref-consent. A no-name / generic product proceeds without a ref.
• AI-creator faces are synthetic personas, not real people. Build a consistent non-existent creator (see workflow). Do not face-mix toward a recognizable real person.
• Music is a separate the selected audio provider pass. Kling --audio is for the VO/lipsync only; ban music in the Kling prompt and overlay an instrumental bed in the editor stage (per MEMORY.md Kling-no-music note).

The niche, in one paragraph

A UGC ad lives or dies on believability and the first 3 seconds. The viewer must feel a real person is talking to them, not a brand. The hook is a problem mirror — show the viewer their own problem so they stop scrolling because they see themselves. Then a transformation/hope beat, then a real verdict, then a low-friction CTA. Trust comes from mannerisms (a nervous laugh, a hair tuck, direct eye contact, pointing at the product) far more than from polished delivery. The single biggest AI tell is a too-perfect, too-symmetrical creator and a studio-clean voice — fight both.

Beat / script shape (default ~15s, 9:16)

Author a shooting script as a table the art-director can fan out from:

Timestamp	Voiceover (exact words)	Visual / shot	Action / mannerism
0-3s	the hook line	problem shown visually, creator to camera	pattern-interrupt gesture
3-7s	the turn	product applied / used / demoed	hands on product
7-12s	the verdict	reaction / before-after	genuine reaction (laugh, raised brows)
12-15s	the CTA	creator to camera	point / "link in bio"

Virality principles to bake in: problem-mirror hook, hope loop (transformation after the problem), audio pattern-interrupt (a whoosh SFX on the cut), mannerisms = trust, room presence (a touch of reverb on the VO so it reads recorded-in-a-room, not studio), CTA urgency (specific, low-friction).

Framing + realism

• Handheld smartphone-UGC feel, eye-level, 9:16. Natural indoor light, not flat phone-flash.
• Photoreal creator. Skin pores, micro-detail, slight asymmetry — fold the anti-AI-slop / photoreal-still register in (pica skill find; see MEMORY.md anti-ai-slop + photoreal-still notes). A flawless face is the giveaway.
• Generate the creator anchor with the product in hand or near the face for the first-frame reference.

Default model stack (verify with pica model search and pica model info)

• Script: the scenarist LLM (the agent's LLM) — feed it max context (product, target pains, the niche's viral patterns, the chosen hook framework).
• Creator + product keyframes: google/gemini-3-pro-image-preview (nano-banana-pro). Generate 3-5 variants, pick the most realistic.
• i2v (multi-shot talking creator): kwaivgi/kling-v3.0-pro — direct it shot-by-shot (each cut = creator speaks → product close-up → reaction → CTA). Kling --audio carries the EN VO + lipsync; for non-EN, VO via the selected audio provider (confirm target language at intake — see MEMORY.md Kling-no-RU-audio).
• Music: separate the selected audio provider Music pass, instrumental, post-mixed in the editor (no artist names — MEMORY.md).
• Captions: per-slot on the locked VO.

Workflow

1. Intake. Collect: product (name / URL / image), target audience + their pain, hook type (problem-solution / before-after / testimonial / transformation — or let the agent pick), target language, hard "no"s. Announce: "This is a UGC ad — using the UGC-ad skill." Defaults: 9:16, ~15s.
2. Reference gate. Named real brand product → get a ref or logged --no-ref-consent. No-name → proceed.
3. Script. Produce the timestamped shooting-script table; get the user's "go".
4. Creator persona anchor (with product), then per-beat anchors — one beat at a time with checkpoints.
5. i2v per beat (Kling, directed shot list), then VO/SFX, then music post-mix, then captions, then the user's chosen composition/render path.
6. Hand off to /evaluator for the post-render quality gate.

Cookbook

• "Make a UGC ad for my no-name vitamin-C serum" → no ref needed; problem-mirror hook (dull skin), creator persona, ~15s 9:16. Match this skill, run the pipeline.
• "Creator ad for the Dyson Airwrap" → named real product → reference gate fires; ask for a ref photo first.
• "Same ad but with a different creator" → NOT this skill → /ugc-model-swap.