RN Human Motion Extractor

Convert a selected range of an authorized reference video into reusable motion data. Preserve motion and timing; do not package the source person's appearance, voice, subtitles, or footage as a reusable identity asset.

Boundary

• Use only a public or local reference the user is authorized to analyze.
• Keep source footage and identity-preserving overlay previews in the private project workspace. The pure-skeleton video and keypoint data are the portable motion assets.
• This Skill extracts evidence; it does not promise that a generative video model will obey every keypoint.
• Do not call a result "exact hand motion" when fingers are occluded, blurred, outside the frame, or below the confidence gate. Report coverage instead.
• Route designed cards, connectors, typography, and UI motion to rn-motion-replica; route full-frame replica claims to rn-replica-qc.

Workflow

1. Lock the range and timebase

Confirm the input file and desired range. Probe the source with ffprobe and record width, height, fps, duration, frame count, and audio presence. Default to the full clip only when the range is already short and unambiguous.

2. Create a local runtime

The bundled extractor requires FFmpeg plus Python 3.10, 3.12. Use an isolated environment; do not modify the system Python.

uv venv --python 3.12 .venv
uv pip install --python .venv/bin/python -r \
  <skill-dir>/scripts/requirements.txt

3. Extract the trajectory

.venv/bin/python <skill-dir>/scripts/extract_motion.py \
  --input <reference.mp4> \
  --output <private-project-dir>/motion-extraction \
  --start 0 \
  --duration 10

Omit --duration to process from --start to the end. Use --no-overlay when an identity-preserving QC preview is unnecessary.

The extractor records: