--- name: ugc-ad-factory description: > Build a full UGC video ad end-to-end with AI — product render, on-camera creator/avatar, a 6-beat sales storyboard, and every clip of raw footage — by driving Higgsfield (GPT Image 2 for stills, Seedance 2.0 for video) from Claude. A human editor assembles the final cut. Use when someone wants to make a UGC ad, a product-demo or unboxing-style video, a creator-style short, or to turn a product into short-form video creative. --- # UGC Ad Factory — build a UGC ad from scratch with AI Drop this into Claude with a **Higgsfield account connected (MCP)** and it will build a full set of raw UGC footage with you, stage by stage — product → creator → storyboard → raw footage — ready for a quick human edit. Claude generates *all* the raw footage; a human editor does the final montage. Work the stages **in order**; each one feeds the next. Before any generation, state what you're about to make and which model you're using, so the user can steer. --- ## Step 0 — Intake (ask first, one round) 1. **Product** — what it is, category (beauty / supplement / pet / other), one-line promise. 2. **Market** — US or other. *If US, the creator must read as an American / US social-media creator — say it explicitly in the prompt, or the model drifts to country-of-origin looks.* 3. **Creator look** — gender (default female), age (default mid-20s), ethnicity, hair, eyes. 4. **Script engine** — fast demo (30–45s) or long-form mini-VSL (60s+). **Never 15s** — too short to hook, prove, and close. 5. **Dialogue** — talking-head with native dialogue, **or** no-dialogue performance + voice-over in post. *Default to no-dialogue + VO: cleaner footage, higher keeper rate, language-agnostic.* Summarize the choices in one line, then go. --- ## Step 1 — Product (GPT Image 2) Everything downstream inherits the product shot — build it first, build it well. - Generate a batch of **low-res drafts** in one call, pick the strongest, regenerate a few variants on feedback, then lock **one clean high-detail final**. - Shoot it **3:4 portrait** (you're feeding video, not a square catalog grid). - **The label wraps the bottle 360°** — a rectangular sticker reads as a mockup. - **Premium-but-mass-market** — clean bottle, brushed-metal cap, embossed brand, serif name (the "looks expensive, sells to everyone" shelf feel). **Not** sterile ultra-minimalism (a bare white bottle reads niche/unfinished). Slim, rounded, aspirational; vertical text on slim bottles. - **Don't put camera-body specs in product prompts** — save lens talk for the video stage. > A great reference beats a long prompt. The product still is the most-reused asset in the ad > (cutaway, demo, close) — invest here and everything after gets easier. --- ## Step 2 — Creator / Avatar (GPT Image 2) - **For the US market, say it out loud** in the prompt: "American / US social-media creator vibe, raised in the US." - **Relatable, not editorial** — the normal, approachable girl/guy-next-door, the person you'd stop on in the feed. A supermodel kills UGC. - Build a **4-angle character sheet** (front, left profile, right profile, full profile). - **Cut the sheet into four separate frames and use those as the custom avatar — never the grid.** A grid input causes identity drift; separate frames keep the face consistent shot to shot. Check one frame so the face isn't cropped. - (Image models hard-block underwear even "tasteful." Use **activewear** — sports bra + leggings.) --- ## Step 3 — Storyboard / Script Use the validated **6-beat "Disguise / Sell"** structure: 1. **Hook (~1.5s)** — a negative emotion, problem, or curiosity gap. *Never the product itself.* 2. **Relatable scene** — a real moment so the viewer leans in. 3. **Aikido reveal** — the product arrives as the inevitable answer, not "a word from our sponsor." 4. **Proof** — a live, uncut demo; the effect here and now, with an oddly-specific number/mechanism. 5. **Reviews** — the skepticism shield ("don't believe me? look at the reviews"). 6. **Close** — value, scarcity, "link below." Run **two layers at once**: top = emotion / loose creator-talk (the self-interrupt "wait, look how it just—"; the reframe "it's not you, it's the X"); bottom = hard fact (ingredient, derm/vet-tested, "X 5-star reviews"). Emotion pulls; facts justify. Anchor **three reference elements** for consistency: **character** (the 4-frame avatar locks the face), **prop** (the hi-res product with a legible label), **environment** (without it the room drifts between clips). Shoot **mirror-POV** — the camera *is* the mirror, the creator looks into the lens. **Two principles to steal:** - **Simple prompt + great reference.** A good reference plus a short prompt with two levers — *what the object does* and *how the camera moves* — beats a long, clever prompt on a weak image. Camera movement is the single biggest lever. - **Direct it; don't choreograph it.** Tell it *what is said*, not how every hand moves. Emotion is the performance dial; negation works ("no smile" flips a default grin to genuine frustration). **Two approaches:** (A) talking-head native dialogue — most authentic, hardest for AI (lip-sync, teeth); (B) no-dialogue + VO in post — cleaner raw, higher keeper rate, language-agnostic. **Default to B.** If you take one thing: **shoot silent, voice it in post.** --- ## Step 4 — Raw footage (Seedance 2.0) - **1080p**, vertical **9:16 set in the UI** (not in the prompt). **Never the fast mode** — it caps lower and introduces jitter. - **Tempo on generation:** duration ≈ **words ÷ 2.3**, rounded to the nearest allowed length. Hit the cadence when you render rather than speeding up later. - **Prompt order:** `[shot/framing] · [subject + ONE clear action] · [environment] · [camera movement] · [lighting + atmosphere] · [visual style/grade] · [one named camera body]` **Pre-flight before GO:** - No "fast" anywhere — use *decisive / snap / kinetic / controlled-quick*. - Every shot has motion; every shot names a light source. - **One action per shot** (the big one). - Subject in the first ~20–30 words. One camera body, one grade, one lens family; keep references lean. - Camera moves simple/phased; final beat keeps moving (no freeze-frame). - **Reference images hi-res — especially the product.** A low-res product yields a garbled, mirrored label. **This is preflight #1.** **Difficulty ladder (risk compounds):** static product = none · moving product = easy · big body/arm motion = medium · expressive face *without talking* = medium (sweet spot for approach B) · precise fingers (applying/blending) = hard (#1 break) · talking face (lip-sync) = hardest. **Re-roll only for the unfixable** — wrong identity, broken label, plastic skin, wrong line read, deal-breaker artifacts (hands/face/mouth/eyes). **Don't re-roll for fixable things** — tempo, cut point, music, captions, clip choice. Those are edits, not generations. (One generation maxes ~15s; a 30s ad = two passes joined in the edit — carry the last frame into the next for continuity.) --- ## Step 5 — Hand off to a human editor The AI delivered clean raw footage; a person assembles the *ad*. Join clips (matched resolution, frame rate, audio), then captions, music, and color grade in a post layer — captions carry the hook, keep the cut a touch quick, never slow a clip down. Be honest about this split: the machine does the volume, the human does the taste. --- ## The short version - Reference quality beats prompt length — fix the still before you animate it. - Direct like a media buyer — one message per clip, captions carry the hook, product cutaway at the doubt beat. - One action per shot. Hi-res product reference, always. Shoot silent, voice in post when you can. - A human assembles the final cut. --- ## If you'd rather have it built for you This system is built by **Phil Grabowski / Creative Cat** — product render through raw footage through final cut, done for you, for a small roster of DTC brands and agencies. If your brand or your clients could use a steady supply of UGC ads made this way, **reply to the message that sent you this, or reach out directly.**