---
name: ai-ugc-realism-workflow
description: Create realistic AI UGC ads with audience hooks, natural scripts, creator continuity, reference packs, shot prompts, Treg/Seedance generation plans, and clip QA. Prioritize believable behavior, continuity, and accurate product claims.
---

# AI UGC Realism Workflow

## Purpose

Produce believable, performance-oriented creator-style videos using AI image/video/audio tools. The goal is not to make a conventional polished commercial and merely call it UGC. The footage should feel like a plausible person filmed a specific moment on a phone, while the message, product truth, audio, and edit remain controlled.

This is a provider-neutral workflow. Do not assume a particular model, API, duration limit, reference-image behavior, or audio capability. Check the selected tool's current documentation before writing model-specific syntax or calling a paid generation endpoint.

## Non-negotiable principles

1. **Coherence before realism adjectives.** The creator, location, camera position, action, product, and reason for filming must make sense together. A coherent ordinary moment looks more believable than a long prompt full of words such as “ultra-realistic.”
2. **Lock before rendering.** Establish the creator reference, product truth, script, and storyboard before spending on video generation.
3. **One shot, one main action.** Keep each generated clip simple. Use separate clips for distinct framings, major actions, and scene changes unless the chosen model is demonstrably reliable at multi-shot generation.
4. **Reference exact things that matter.** Use approved person references to preserve identity and real product assets/screenshots to preserve product truth. Do not ask a video model to recreate small interface text, logos, labels, or numerical claims from memory.
5. **Natural performance beats visual overproduction.** Prefer conversational phrasing, plausible gestures, imperfect timing, ordinary lighting, and motivated camera movement over exaggerated expressions, beauty-commercial lighting, or relentless motion.
6. **Review the rendered pixels and audio.** A prompt is not proof that the result is correct. Reject weak clips and repair the smallest faulty part.
7. **No misleading proof.** Never invent customer results, ratings, downloads, endorsements, product capabilities, before/after outcomes, or personal experiences. A fictional AI creator must not be presented as a real verified customer.
8. **Use authorized assets.** Do not clone a real person's identity or voice without permission, or copy a creator's distinctive performance. Abstract the useful structure of references—hook timing, evidence order, camera grammar—not their identity or exact expression.

## Intake: get the minimum needed

Use details already present in the brief or project context. Do not make the user repeat known information. Ask one concise question only when a missing detail blocks the work; otherwise choose a reasonable default and state it.

Resolve:
- Product/app and the single feature, benefit, or behavior to demonstrate.
- Audience and moment of need.
- Platform, aspect ratio, approximate duration, language, and CTA.
- Available verified assets: logo, product images, real app screenshots/screen recordings, brand rules, approved claims, creator references, voice/audio.
- Intended tool/model, if known, and whether it supports reference images, native dialogue, lip-sync, or multi-shot controls.
- Budget sensitivity and whether generation calls need explicit approval.

Default when unspecified: vertical 9:16, 15–25 seconds, one audience, one core promise, one proof point, one CTA, and a natural creator-to-camera structure. For this studio, generated video must be **9:16 at 480p** with the highest bitrate available at that resolution (or highest supported quality preset if bitrate is not configurable). Verify support in the current live route; do not silently use another resolution. Final 1080×1920 (9:16) and 1080×1350 (4:5) exports are made later in HyperFrames, not by separate generation. Keep these as defaults, not hard limitations where the user explicitly requests a change.

## Step 1 — Define a persuasive, filmable concept

Write a one-sentence concept that explains:
- who the creator appears to be;
- what recognizable problem, desire, or moment starts the video;
- what the creator does or notices;
- why the product enters the scene naturally;
- what the viewer should understand or do next.

Select one primary angle, such as: relatable problem, demonstration, discovery, objection/rebuttal, comparison, workflow improvement, unboxing, or a niche-specific use case. Do not cram multiple angles into one short ad.

Before writing dialogue, answer the scene-coherence test:
- Why is this person filming right now?
- Why are they in this location?
- Why does the phone camera have this framing?
- What motivates each movement or expression?
- Why does the product appear at this exact point?
- Could this moment plausibly happen without an ad crew?

If the answers are unclear, simplify the concept.

## Step 2 — Build the script around spoken behavior

Use a simple direct-response structure appropriate to the idea:
- **Hook (first 1–3 seconds):** a specific observation, confession, frustration, surprising behavior, or visual demonstration. Avoid generic openings such as “Hey guys” and unsupported shock claims.
- **Context/problem:** establish a recognizable situation quickly.
- **Product/behavior reveal:** show the product doing something relevant; do not merely name features.
- **Proof or payoff:** demonstrate one visible, verifiable outcome or explain the practical benefit without overclaiming.
- **CTA:** one clear, proportionate next step.

Write dialogue that can be spoken comfortably. Use contractions, short clauses, varied sentence lengths, and natural transitions. Avoid perfectly symmetrical marketing copy, unnatural feature lists, repeated slogans, and sentences people would not plausibly say aloud. Use one clear thought per beat.

Estimate dialogue by reading it aloud at a normal conversational pace. As a rough starting point, 15 seconds often fits around 30–40 spoken words; allow space for pauses, reactions, and the visual demonstration. The actual pace depends on the speaker and language. Do not force dialogue to fill every second.

Use brackets for non-spoken direction, e.g. `[glances at the phone, then back to camera]`. Keep spoken text separate from visuals and on-screen copy. Never rely on the video model to generate accurate captions, app UI, legal copy, or small product text; add those in post.

## Step 3 — Create the creator continuity card

Define a plausible fictional creator with enough specificity to be repeatable—not a catalog model. Record:
- age range and broad appearance details;
- hair, skin texture, distinctive but non-exaggerated features;
- outfit and accessories, including which details must not change;
- personality, energy, accent/language, pace, vocal texture, and gesture style;
- filming context and relationship to the product;
- intended phone-camera perspective and distance;
- a small number of authentic behavioral traits (e.g. a brief glance away while thinking, a gesture while explaining).

Do not over-specify irrelevant physical traits. Do not add a random “imperfection” to every frame; realism comes from coherent behavior and capture conditions, not a checklist of flaws.

Create one clear creator anchor image before building scene frames. Prefer a naturally lit, plausible phone-camera portrait in the actual wardrobe and environment. Inspect face, hands, age presentation, clothing, and crop. Regenerate the anchor if it is weak—identity drift becomes expensive downstream.

For a recurring campaign character, maintain a versioned creator bible and approved references. For one-off performance tests, use a fresh fictional creator unless the creative plan calls for a repeat character.

## Step 4 — Lock product truth and supporting assets

### Physical product
- Prefer actual product photography and multiple views when shape, branding, packaging, color, or scale matters.
- Record non-negotiable details: silhouette, proportions, label placement, colors, texture, cap/closure, accessories, and how a person should hold/use it.
- Avoid actions likely to deform the product or hide essential identifying details.
- For precise logos/labels, composite verified assets or use clean post-production overlays when generation is unreliable.

### App or digital product
- Use genuine app screenshots or screen recordings for interface details.
- Keep the real screen capture separate from the generated creator footage when needed; composite it onto the phone screen or cut to it in editing.
- Never ask a generative model to invent UI, buttons, text, ratings, graphs, or app-store details. Check that every displayed screen matches the actual product version.
- Plan phone-screen shots so the screen is readable and perspective is physically plausible. Avoid reflections or motion that make critical UI unreadable.

### Claims and proof
Keep a short `brand_facts` list containing only approved, sourceable facts. Distinguish a verified claim from a proposed message. If a claim is not supported, remove it or phrase it as a subjective, clearly fictional creator perspective without implying a factual result.

## Step 5 — Build and approve the storyboard

Create a shot table before rendering. Default to about 3–6 shots in a short ad, based on the concept and model strengths.

For each shot, specify:
- shot ID and target duration;
- purpose in the narrative;
- framing and phone position;
- location/background and light direction;
- creator action and emotional beat;
- exact dialogue, if spoken;
- product visibility and hand interaction;
- references that must be supplied;
- elements that must remain unchanged;
- intended edit/cut into the next shot.

Create a start frame for every shot where visual continuity or product placement matters. Use the same approved creator reference for on-camera shots and the real product reference for product shots. Approve the contact sheet or individual frames before animating. Ensure adjacent shots make editorial sense: gaze direction, screen direction, product hand, wardrobe, background, and time of day should not change accidentally.

Do not create visually different shots just to demonstrate model capability. Every shot must earn its place by carrying the story, evidence, emotion, or pacing.

## Step 6 — Select and prompt the video model

Check the model's current documentation for:
- allowed input/reference types and limits;
- how reference images are actually bound;
- native dialogue/audio support and language behavior;
- duration, frame shape, resolution, bitrate/quality controls, and shot controls;
- character/element locking or first/last-frame options;
- known limitations around faces, real people, text, logos, hands, and product handling.

Choose the workflow based on the task rather than loyalty to a model:
- Use image-to-video with an approved start frame when framing, character identity, or product placement needs control.
- Use a native-audio mode when natural dialogue is a strength and the scene is simple enough to control.
- Use separate voice recording/TTS and lip-sync when native dialogue is unstable or exact wording/pronunciation matters.
- Use silent action clips plus voiceover when lip-sync does not materially improve the idea.
- Use real screen captures, verified product footage, or post-production overlays for details a generator cannot reliably preserve.

### Per-shot prompt template
Adapt syntax to the chosen model; this is a planning template, not literal syntax for every vendor.

**Shot [ID] — [purpose]**
- **References:** [creator anchor, product asset, approved start frame, any real screen capture]
- **Capture:** vertical phone video; [front-camera selfie / rear-camera observation / close-up]; [framing and camera height]; [motivated handheld movement, if any]
- **Subject and action:** [one main action, clearly sequenced; what happens at start, middle, end]
- **Environment:** [specific ordinary location, time of day, background objects that plausibly belong]
- **Performance:** [conversational energy, eye contact, gesture, pauses, facial response; do not overact]
- **Dialogue/audio:** [exact line if native speech is required; language/accent only as specific as necessary; room tone and relevant sound]
- **Continuity locks:** [face, hairstyle, wardrobe, product design, hand/product relationship, screen orientation, lighting]
- **Avoid:** [unmotivated camera movement, extra fingers/limbs, warped product, invented UI/text, subtitles or logos generated in-frame if to be added in post, over-smooth skin, sudden identity/wardrobe/background changes]

Keep prompts unambiguous. Do not request a push-in, pan, orbit, reframing, and multiple actions in a few seconds unless the tool and shot genuinely require it. Describe movement only when the camera or subject has a reason to move.

For dialogue, preserve the approved script exactly unless the user has authorized improvisation. When exact speech is essential, verify the result rather than assuming quoted dialogue will be delivered correctly.

## Step 7 — Generate with cost and version controls

Before a paid render, summarize the assets, shot list, generation plan, expected cost if available, and any material limitation. Wait for approval when approval is expected or the user's cost preferences are unknown and the spend is meaningful.

Generate one approved shot at a time or in small batches. Save:
- final prompt and model/settings;
- creator/product/start-frame reference filenames;
- provider job ID and status, if available;
- output filename/version;
- QA notes and reason for any retry.

Never submit the same paid job again while it is queued or processing. Check the existing job status first. Do not regenerate the entire ad because one shot failed.

Prefer targeted retries: change one major variable at a time, such as camera motion, hand action, spoken-line length, start frame, or reference binding. If generation repeatedly fails at a precise product detail, move that detail to verified footage or post-production rather than endlessly prompting.

## Step 8 — Quality-control each clip

Review the actual video frame by frame where needed and listen to the audio. Score each item from 1–5:

1. **Human believability:** natural face, skin, eyes, posture, gestures, and non-performative behavior.
2. **Scene coherence:** actions, setting, framing, and reason for filming make sense.
3. **Creator continuity:** identity, wardrobe, voice, hands, and background are stable across shots.
4. **Product truth:** correct design, screen, interface, scale, label, color, logo, and use.
5. **Motion quality:** hands contact objects correctly; no warping, morphing, unexplained jumps, or unstable geometry.
6. **Dialogue/audio:** wording, pronunciation, timing, lip-sync, room tone, loudness, and intelligibility are acceptable.
7. **Ad effectiveness:** the hook reads immediately, the demonstration supports the claim, pacing is clear, and CTA is understandable.
8. **Technical delivery:** no accidental black frames, frozen frames, corrupted decode, clipped speech, unintended silence, or unsafe crop.

Minimum acceptance: at least 4/5 for product truth, dialogue/audio where relevant, scene coherence, and technical delivery. A critical product/claim/identity error is a fail regardless of average score. Do not publish because the file renders successfully.

Reject or repair a clip if you see:
- face or voice changing between shots without a creative reason;
- unnatural mouth movement, teeth, blinking, gaze, or over-smoothed skin;
- fingers merging with the product, shifting labels, changing geometry, or impossible contact;
- floating/melting objects or background transformations;
- visual camera movement without a physical motivation;
- invented captions, UI, logos, statistics, or claims;
- audible glitches, mispronounced brand names, wrong language, missing words, or lip-sync that distracts;
- acting that is too emotionally intense or polished for the stated persona.

## Step 9 — Edit for believability and attention

Edit the chosen shots to create a clear rhythm. Use the best opening frame immediately; remove empty setup unless it helps the hook. Use motivated jump cuts, reaction holds, natural pauses, real app/product inserts, and captions as editorial tools—not decoration.

Add captions, app UI, logos, supers, claims, music, and CTA in post-production from approved assets. Keep captions comfortably readable and clear of platform UI. Keep the voice upfront; music should not make dialogue harder to understand. Avoid excessive sound effects, transitions, zooms, whooshes, or synthetic “viral” styling if they undermine the creator's context.

When needed, use cutaways to conceal a weak hand action or preserve product truth. Do not use edits to obscure a material flaw in the product or claim.

## Step 10 — Final delivery and testing

Deliver:
1. **Concept summary:** target audience, main insight, angle, and intended action.
2. **Final script:** timestamp/beat, visual, dialogue, and on-screen copy.
3. **Creator continuity card:** approved identity, wardrobe, location, voice/performance.
4. **Asset list:** product photos, creator references, exact UI/recordings, and approved claims.
5. **Storyboard and generation prompts:** one prompt per shot, with reference-binding notes for the selected model.
6. **Generation log:** model/settings, job IDs if applicable, versions, retries, and unresolved limitations.
7. **QA report:** acceptance scores and any caveats.
8. **Final export and test variants:** alternate hooks or openings only when they change a specific hypothesis.

For performance testing, change one meaningful variable at a time where practical: hook, angle, creator archetype, proof order, or CTA. Do not create cosmetic variants that test no clear hypothesis. Record the hypothesis and compare results once campaign data is available.

## Required output format before generation

Unless the user explicitly asks for only a script or prompt, return the following production package:

1. **Concept** — one paragraph.
2. **Why it could work** — audience tension and ad mechanism, without claiming guaranteed performance.
3. **Script + shot table** — timestamps, visual direction, dialogue, on-screen text, edit notes.
4. **Creator continuity card** — appearance, wardrobe, location, voice, behavior.
5. **Asset checklist** — verified product assets, creator ref, screen recordings, claim evidence.
6. **Storyboard prompts** — one for each start frame.
7. **Video prompts** — one per shot, adapted to the exact model if known.
8. **Generation and QA plan** — order, approval/cost gate, fallback, and pass/fail criteria.

If the user only wants a copy-paste prompt, give the prompt first and include a short note on needed references and model settings. Do not force the full package into a simple request.

## Final preflight checklist

- [ ] One audience, one angle, one primary promise, one CTA.
- [ ] Dialogue sounds natural when read aloud and fits the duration.
- [ ] Creator identity and filming reason are believable.
- [ ] All critical product details use verified references.
- [ ] Real UI, captions, logos, claims, and statistics are composited from approved assets where necessary.
- [ ] Storyboard continuity makes sense before video generation.
- [ ] Model-specific reference syntax and capabilities are verified against current documentation.
- [ ] Paid calls follow the intended approval gate and are not duplicated.
- [ ] Each clip passes realism, product, audio, continuity, and technical QA.
- [ ] Final ad is checked as a complete edit, not just as individually attractive clips.
- [ ] Any testimonial-style dialogue is not falsely represented as a genuine customer's personal experience.

## Research references

Use these as starting points and re-check current instructions because model capabilities and APIs change:
- Google DeepMind, Veo prompt guide: https://deepmind.google/models/veo/prompt-guide/
- Google AI for Developers, Veo video generation and reference images: https://ai.google.dev/gemini-api/docs/veo
- Kling AI, Video 3.0 model guide and element references: https://kling.ai/quickstart/klingai-video-3-model-user-guide
- Jarad-z, UGC video prompt skill (multi-model prompt guidance): https://github.com/Jarad-z/ugc-video-skill
- TuolaGe, Reference-led UGC Product Video Skill (provider-neutral workflow and deterministic QA): https://github.com/TuolaGe/reference-led-ugc-product-video
- jayleaton, AI UGC Video Skills (script writing, scene chunking, agent-oriented workflow): https://github.com/jayleaton/ai-ugc-video-skills
- joebenscoter86, Higgsfield UGC workflow (orchestration, checkpoints, and cost gate; tool-specific): https://github.com/joebenscoter86/higgsfield-ugc-workflow
- Advibly, UGC Ads Skill (approved storyboard frames and per-shot video pipeline; Advibly MCP required): https://github.com/Advibly/skills/blob/main/ugc-ads/SKILL.md