---
tier: hero
reliability: 4.0  # reference-sheet gen 0 + ltx-q-ingredient .6 + music .2 + ffmpeg 0 → 3.2; +0.8 standard public caps
---

# Reference Sheet → Creative Ad — agent playbook ⭐

> Paste this whole file (with the BRIEF filled in) into Claude cowork, chat, or Code. **This is the LTX 2.3 Ingredient showcase.** One reference sheet — your character, product, and props on a single image — becomes a generated video scene that composites them together, on-model. The pipeline that turns "here are my ingredients, put them in a scene" into a finished ad.

## What you'll get

- **One reference sheet** — a single image laying out the ingredients (hero product, character, props, an environment swatch), built for you
- **A video scene** via **`ltx-q-ingredient`** (LTX 2.3 Ingredient) that brings those ingredients together in the action you describe
- **A scored, platform-ready cut** — `music` / `sonilo-v2m` bed + `ffmpeg-export` to aspect + optional `ffmpeg-overlay` logo
- Multi-scene option: 3–5 ingredient scenes from the same sheet → `ffmpeg-concat` → one ad
- Total: ~3–5 min for a single scene, ~$0.40–$1.00

## The BRIEF (fill this in)

```
INGREDIENTS: <list them — e.g. "hero sneaker (the product), a young runner (character), a city-rooftop at dusk (environment), a gym bag (prop)">
PRODUCT IMAGE(S): <public URL(s) of the real product, if you have them — host local files via scripts/host-local-file.sh>
SCENE: <what happens — e.g. "the runner laces up on the rooftop, the sneaker catches the last light, slow push-in">
STYLE: <look / mood / camera — e.g. "warm cinematic, shallow depth, golden hour">
ASPECT: <9:16 | 16:9 | 1:1>
MUSIC: <vibe, or 'auto from the video' for sonilo-v2m>
```

## The pipeline (stages)

1. **Build the reference sheet** (the key step). Lay all ingredients into ONE clean image:
   - If you have a real product photo, composite it with generated character/props/environment cells using `pillow-grid`, or
   - Generate the whole sheet with `create_media(action:"generate", model_override:"gpt-image")` (or `nano-banana` / `flux-dev`): *"a clean reference sheet / contact grid: cell 1 the hero sneaker on white, cell 2 the runner full-body, cell 3 a rooftop-at-dusk swatch, cell 4 a gym bag — neutral backgrounds, consistent scale, labeled."*
   - Keep each ingredient **unambiguous** (clear silhouette, neutral backdrop) so the model can tell product from prop.
2. **Ingredient → video** — `create_media(action:"animate", source_url:<reference sheet>, model_override:"ltx-q-ingredient", prompt:"<SCENE + STYLE>")`. The sheet locks *what appears*; the prompt directs *what happens*.
3. **(Optional) more scenes** — re-run stage 2 with different SCENE prompts on the same sheet for a multi-shot ad; `ffmpeg-concat` them.
4. **Finish** — add a soundtrack (`music`, or `sonilo-v2m` to auto-score from the video) + `mirelo-sfx-v2v` for foley; `ffmpeg-export` to aspect; `ffmpeg-overlay` the logo. (See the `finishing-quality` skill.)

## Why Ingredient here

`ltx-q-ingredient` is for when **the identity of the things in the shot matters** — known characters, a specific product, branded props — not generic text-to-video. It's the video analog of moodboard-to-scene. See the `ltx-ingredient` skill for reference-sheet craft.

## Tips
- **The sheet is everything** — coherence of the output tracks the clarity of the sheet. Invest the first minute there.
- **Separate *what* from *what happens*** — ingredients in the sheet; action + camera + mood in the prompt.
- **Brand-lock it** — pair with a `brand_kit` / `context_kit` so the palette + product carry across every scene.