Pro

H3 Video Reverse

Turn a reference video into a shot breakdown and original MiniMax H3 prompts

$59.90

For creators and marketers who see a video and think "I want one like that" — made with MiniMax H3 (Hailuo 3).

Give the skill a video file (or describe one) and it works like a director's assistant:

  1. Extracts evidence. A bundled Python script uses ffmpeg to detect scene cuts and export keyframes from the start, middle and end of each shot, a contact sheet stamped with frame numbers and timestamps, and JSON metadata: duration, fps, resolution, aspect ratio, whether there is audio, plus loudness and silence data. Standard library only; it works on macOS, Windows and Linux.
  2. Breaks the video down shot by shot — shot size, lens feel, camera move, action, timing, lighting, color, sound design, music, transitions and end state — with every judgment labeled observed, inferred or unknown.
  3. Distills the style DNA (pacing, framing, camera language, grade, sound strategy) and swaps in your own subject, new dialogue, an original music description and your own branding.
  4. Writes the H3 prompts, split into 4–15 second clips at shot boundaries, in the right format for the Hailuo app, the API or self-hosted ComfyUI/SGLang.

The skill cannot hear audio: it reports loudness, asks you a few targeted questions about speech and music, and marks anything it assumes. No ffmpeg or no file? A described-video workflow builds the same breakdown from your notes or screenshots. Requires Python 3.9+ and ffmpeg for file analysis.

Highlights

  • Script finds scene cuts, keyframes, a labeled contact sheet and metadata
  • Shot table: size, lens, move, action, light, color, sound, music, transitions
  • Rebuilds the style around your subject; replaces dialogue, music and logos
  • Splits long or fast-cut videos into H3-sized clips with chaining notes
  • Honest about audio: loudness data plus targeted questions, never guesses

Example

Analyze an ad file
You

Here's a 13-second bottle ad I love: ./bottle_ad.mp4. Make the same kind of video for our green insulated tumbler. I use the Hailuo app.

Agent + skill
## Source
bottle_ad.mp4 · 13.0 s · 1080×1920 (9:16) · 30 fps · audio: yes · 5 shots (avg 2.6 s, fast) · cuts at 1.4 / 3.1 / 6.0 / 9.2 s
Audio (loudness only): silence 5.6–6.0 s, level jump +14 LU on the 6.0 s cut. I can't hear the track — is that a music drop and an impact? Any voice?

## Shot breakdown (excerpt)
| # | Time | Size | Lens | Move | Action | Light | Out | Conf. |
|---|---|---|---|---|---|---|---|---|
| 1 | 0–1.4 | ECU | macro, razor DOF | push in, small, slow | droplets slide down a matte bottle | cool backlight | cut | H |
| 4 | 6.0–9.2 | low-angle CU | wide | arc, slow-mo | bottle lands on wet concrete, splash | hard backlight | cut | H |

## Style DNA
Macro tease → tactile action → human context → hero drop on a hard hit → packshot end card · silent beat before the impact · teal shadows, warm highlights, light grain.

## H3 plan
1 clip · 13 s · 9:16 · Omni Reference (your product photos as Image 1–2) · source shots 1–2 merged into one macro push-in.
No ffmpeg: work from a description
You

I can't install ffmpeg at work. It's an 8-second anime fight: wide rooftop in rain, super close sword clash with sparks, then one lands behind the other in slow-mo. Rain, clash, a shout, taiko. I run H3 in ComfyUI.

Agent + skill
Working from your description (confidence mostly medium).

| # | Time (est.) | Size | Move | Action | Sound | Music |
|---|---|---|---|---|---|---|
| 1 | 0–3 | EWS | slow push (assumed) | two fighters face off on a rainy rooftop | rain | taiko starts |
| 2 | 3–5 | ECU | shake strongly | blades clash, sparks | clash + shout | hits |
| 3 | 5–8 | low-angle MS | slow-mo arc (assumed) | lands behind rival, coat flaps | rain, thud | resolves |

New characters: a courier with a prosthetic arm vs. a masked rival in a split-tail coat.

```text
integrated_multimodal_description: [Shot 1] 2D-animated, cel-shaded anime style with heavy rain streaks and neon reflections. An extreme wide shot frames a flat rooftop at night… [Shot 2] At 00:03.000, the camera cuts to an extreme close-up as the two blades collide… the courier (S1) shouts, <d>[English] Move!</d> [Shot 3] At 00:05.000, …

overall_soundscape: Heavy rain hammers the rooftop…

non_diegetic_music: Taiko drums at a fast tempo…
```

What's inside

h3-video-reverse/
├── agents/
│   └── openai.yaml
├── references/
│   ├── audio-without-ears.md
│   ├── breakdown-template.md
│   ├── described-video-workflow.md
│   ├── examples.md
│   ├── h3-translation.md
│   ├── originality-guardrails.md
│   └── reading-frames.md
├── scripts/
│   └── extract_frames.py
├── LICENSE.txt
├── README.md
└── SKILL.md

Install by unzipping into your agent's skills folder. Install guide →