OMNI-UGC-AD-FACTORY

# Omni UGC Ad Factory A Claude Code skill that turns a product and an actor reference photo into a finished **~20-second UGC talking-head ad**: two chained 10-second Gemini Omni clips that play as one unbroken take, shipped as a single seamless 9:16 MP4. Built on the [MaxFusion](https://maxfusion.ai) MCP. Real outputs of this pipeline (click a preview for the full MP4 with sound): | | | | |:---:|:---:|:---:| | [](examples/final_916_gym.mp4) | [](examples/final_916_room.mp4) | [](examples/final_916_room_alt.mp4) | ## Why it exists Anyone can prompt a talking head. The hard parts are everything that makes viewers believe it: an actor that looks captured rather than rendered, a script that sounds like a person and not an ad, a voice that isn't monotone, micro-behaviors (uneven pauses, a gaze drop, a nose exhale) that land on the right beats, and two clips that join without a visible seam. This skill encodes all of that as **machine-enforced gates**, not prose advice: ``` casting.json -> gate_casting.py -> approved_casting.txt -> gpt-image-2 clip1.json -> gate_omni.py -> approved_clip1.txt -> gemini-omni-flash clip2.json -> gate_omni.py -> approved_clip2.txt -> gemini-omni-flash ``` The agent never hand-writes a generation prompt. It writes structured JSON; the gate validates it against data-file glossaries (age-correct skin physiology, a 32-entry micro-behavior glossary keyed to internal states, a six-category voice spec, a camera color-grading vocabulary) and *assembles* the prompt itself on PASS. A prompt that skipped its gate does not exist. ## The chaining trick Omni holds its reference image for the first frames of a clip before dissolving into the generated scene. Seed clip 2 with clip 1's trimmed final frame and that quirk becomes an invisible join — measured seams of ~3/255 against a hard gate of 5/255. `scripts/adkit.py` measures speech, trims dead air in the right order, verifies the seam, and stitches the deliverable. ## Install Drop the `omni-ugc-ad-factory/` folder into your Claude Code skills directory (`~/.claude/skills/` or a project's `.claude/skills/`). Requirements: - Claude Code with the MaxFusion MCP connected - `ffmpeg` on PATH, Python 3 with `numpy` and `Pillow` (for `adkit.py`) Then ask for an ad: *"make me a UGC ad for <product>"*. The skill asks five things once (actor age bracket, delivery tone, visual lane, research leads, an actor reference photo) and runs the whole pipeline: competitor teardown, gated script, gated casting, two chained clips, seam check, stitched 9:16. ## Layout - `SKILL.md` — the pipeline, step by step, with every production failure baked in - `data/` — the four glossaries the gates enforce (edit these to improve the system) - `scripts/` — `gate_casting.py`, `gate_omni.py`, `adkit.py` - `references/` — deep references: pitfalls, Omni beat-prompting, casting realism + age-texture decoupling, delivery (how a real person talks), humanizer - `examples/` — finished 9:16 ads produced by this pipeline ## Credits - `references/humanizer.md` is [blader/humanizer](https://github.com/blader/humanizer) (MIT, by Siqi Chen), applied here to ad scripts. - Video generation, image generation, ad research, and canvas analysis run on the [MaxFusion](https://maxfusion.ai) platform via MCP. MIT licensed.