oliverhees/hermes-motion-pack

Ingests the user's own footage (clips, gameplay, events, vlogs, screen recordings) for a highlight video — makes proxies, detects scene cuts, scores every second by motion, loudness and speech, transcribes, and writes SHOTLIST.json with the best candidate moments. Use when the user hands over video files/folders, says "aus meinem Material", "Footage", "beste Momente", "Highlights rausziehen", "schneid mir das zusammen", or before mp-highlight when real footage is the basis.

hermes-motion-pack とは？

hermes-motion-pack is a Claude Code agent skill that ingests the user's own footage (clips, gameplay, events, vlogs, screen recordings) for a highlight video — makes proxies, detects scene cuts, scores every second by motion, loudness and speech, transcribes, and writes SHOTLIST.json with the best candidate moments. Use when the user hands over video files/folders, says "aus meinem Material", "Footage", "beste Momente", "Highlights rausziehen", "schneid mir das zusammen", or before mp-highlight when real footage is the basis.

対応~Claude Code~Codex CLI~Cursor

npx skills add https://github.com/oliverhees/hermes-motion-pack/tree/HEAD/skills/mp-footage

ドキュメント

mp-footage — find the moments

Goal: turn hours of material into a ranked SHOTLIST.json the editor (mp-highlight) can cut from. Never watch everything frame by frame; let the scripts rank, then look only at the candidates.

Steps

Inventory: bash <skill-dir>/scripts/inventory.sh <folder> → footage/INVENTORY.json (duration, fps, resolution, audio) and 720p proxies in footage/proxy/ (edit on proxies, render from originals).
Scan: python3 <skill-dir>/scripts/scan.py footage/proxy/<file>.mp4 --out footage/scan_<name>.json --intent action|emotion|recap → scene cuts + per-second scores (motion, loudness, speech-likelihood) + top moments.
Transcribe if speech matters: npx hyperframes transcribe <file> (or whisper if installed) → footage/<name>.words.json. Quotes with timestamps become title candidates.
Candidates: merge scans into SHOTLIST.json (template below). Rank by the brief's intent (action → motion+loudness; emotion → speech+faces; recap → variety across time). Target 3× the material the final cut needs.
Look at candidates: bash <skill-dir>/scripts/peek.sh <file> <t1> <t2> ... → one 12-frame strip per candidate. Keep the ones that read instantly, drop the rest. Ask the user only when two candidates are equal.
Hand off to mp-highlight with SHOTLIST.json.

`SHOTLIST.json`

{ "source": "footage/", "intent": "action", "target_duration_s": 30,
  "shots": [
    { "id": "A01", "file": "gameplay_01.mp4", "in": 83.4, "out": 86.1, "score": 0.91,
      "why": "peak motion + crowd loudness", "tags": ["action","wide"], "quote": null, "crop_x": 0.5 },
    { "id": "A02", "file": "vlog_03.mp4", "in": 412.0, "out": 414.6, "score": 0.84,
      "why": "laugh + close-up", "tags": ["emotion","close"], "quote": "das war verrückt", "crop_x": 0.62 }
  ] }

Rules

Shots are 0.6–3.0 s for action, 2–5 s for emotion, never longer than 6 s in a highlight.
Keep a 0.3 s handle on both sides for the editor.
Variety: alternate wide/close, loud/quiet, fast/slow. Flag it in tags.
Vertical delivery from horizontal footage → note the subject position per shot (crop_x 0–1) so the editor reframes instead of centre-cropping.
Never upload the user's footage anywhere; everything stays local.

references/scoring.md explains the scores and how to tune thresholds per material type.

Files in this skill

references/scoring.md
scripts/inventory.sh
scripts/peek.sh
scripts/scan.py