ToXMon/ugc-ad-creation

UGC-Ad Creation

A UGC avatar ad with motion graphics costs about $5 — $4.75 projected for the video, plus a couple of cents for motion graphics — compared with $300–$500 from humans for a single ad. Schneider expects ~$3 with another provider. Ten ads = $50; a hundred = $500. That cost gap buys velocity — many variants tested, fast kill/scale decisions — not a replacement for a real customer's voice. Sometimes the AI avatars even outperform humans reading the same script.

The pipeline

This is the pipeline as described on the episode, step by step. Don't skip steps — each one exists because the naive version fails (invented enunciation, plasticky avatars, garbled words).

1. Script — from the static-ad concepts (/UGC, "write a script for this based on above"). Hook (first 3 seconds), problem, product as the fix, proof, CTA. One pain per video. Write for speaking, not reading.
2. Audio first — the script is read with a FAL voice clone → an MP3. Seedance is not given the script alone: left to itself it invents enunciation.
3. Walled garden — the MP3 + a reference image of the creator + the exact script go to Seedance together, so the avatar, voice, and words are constrained to what you supplied.
4. V1 is 480p — 1080p costs ~$14; 480p + an ffmpeg upscale + blur gets you there cheaper. Render takes ~5 minutes. The whole workflow runs in a hosted cloud harness, not on a local machine.
5. Word-by-word verification — transcription is matched against the script to confirm an exact read. When it mispronounces a word (the episode's example: "Kalani" coming out as "Kalandla"), the transcript gives the timestamps and the agent swaps in the correct pronunciation.
6. Source image — the avatar starts from a real creator's image, remixed (background, age, ethnicity changed) — explicitly not a likeness — which kills the plasticky Seedance look.
7. Captions + motion graphics — Mirage Tesseract takes the word-by-word transcription and adds captions, motion graphics, popups, and the call to action.
8. Creative-director gate — a vision model goes through the video frame by frame for alignment before anything goes live; below 8/10 it loops back for fixes (skills/creative-direction-eval/SKILL.md).

The avatar can also hold the product: provide product reference images and say so in the prompt. There are hundreds of avatar variants by demographic and psychographic.

Full prompt: prompts/ugc-script.md.

Human creators

Human creators layer in through Tribe and Sideshift APIs — $40 a video in the episode's example, per-video pay plus a bonus based on spend. The script goes to creators via the API (they may modify it); the agent is the creator manager: it briefs, collects, scores, and scales human-made UGC the same way it scores AI-made UGC. Human-made cuts get the same motion-graphics treatment. Use humans when the claim needs a real person (testimonial, usage demo, regulated claim) and AI avatars when the variable under test is the hook, angle, or offer.

Rules

• Disclosure. AI-avatar / synthetic-creator ads are labelled as AI-made where the platform requires it, and never presented as a real customer's words. A synthetic testimonial is a fabrication — never write one.
• One pain per video. A script that tries to cover three pains covers none.
• Hook in 3 seconds. If the first line doesn't restate the viewer's pain in their words, rewrite it.
• Captions always on. Assume muted autoplay.
• Same scoring as static. Every UGC variant goes through the creative director (8/10 gate) and the experiment log. A $5 ad that nobody watches is still wasted spend — just cheaper waste.
• Expect failures. About 1 in 50 Seedance generations is garbled nonsense. At this cost, discard and regenerate — don't engineer around it.
• Cost honesty. The ~$5 figure is the episode's stated cost ($4.75 video + cents for motion graphics); verify current tool pricing before budgeting. Human-creator pricing ($40/video + spend bonus) is likewise stated, not verified.