convez.ai
SkillAutomaçãoIntermediário
nateober

Ceremony Clip

nateober/ceremony-clip

Cria clipes de cerimônia em 35s com música gerada localmente

Ver no GitHub
InicialPluginScore · de 0 a 100

Skill para Claude Code que transforma gravações longas de cerimônias em clipes cinematográficos de ~35 segundos. Detecta automaticamente o momento em que o nome é anunciado, adiciona cartela de título e compõe uma trilha sonora orquestral original gerada localmente, tudo rodando offline em Mac com Apple Silicon.

Instalação

$npx convez-ai nateober/ceremony-clip

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
1
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 1 mês
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Extrair momento de formatura/cerimônia de arquivo de 2+ horas
  • Gerar trilha sonora orquestral original com MusicGen
  • Criar vídeo compartilhável com cartela de título e áudio verificado

Veredito

Skill bem executado que resolve um problema real: transformar vídeos longos e chatos em momentos memoráveis. A verifikação do próprio trabalho (re-transcrição da mistura final) é engenharia sólida, embora o requisito de Apple Silicon limita a audiência potencial.

README

A tradução para o português deste README ainda está na fila.

ceremony-clip

Turn a multi-hour ceremony recording into a ~35-second cinematic highlight clip of your kid's moment — their name being called, the walk across the stage — with a title card and an original orchestral score, generated entirely on your Mac. No video editor, no stock-music licensing, nothing uploaded anywhere.

Built as an Agent Skill for Claude Code: you hand Claude the recording and a name; it finds the moment, cuts it, scores it, mixes it, and verifies its own work (it re-transcribes the final mix to prove the name is still audible through the music).

What you get

  • The right ~40-second window found automatically — no scrubbing through 2+ hours
  • The right kid verified on screen — at big ceremonies walkers lag the announcer by ~20 seconds, so the graduate crossing during the name call is someone else; the skill frame-checks the window and asks you to confirm
  • Fade-in title card (name + school + honors)
  • A unique orchestral cue composed locally by MusicGen, ducked under the announcer so the name rings clear, bowing out to the real applause
  • A small shareable MP4 your family will actually watch

Requirements

  • Apple Silicon Mac (M1 or later; 16 GB RAM is fine)
  • Claude Code
  • ffmpeg (brew install ffmpeg) and uv (brew install uv)

Python dependencies (Whisper, Silero VAD, MusicGen) install themselves on first run via uv. First run downloads ~3 GB of models; after that it's all offline.

Use it

Install as a skill:

mkdir -p ~/.claude/skills && git clone https://github.com/nateober/ceremony-clip ~/.claude/skills/ceremony-clip

Then, in Claude Code:

Use the ceremony-clip skill to make a highlight clip of Jordan Alvarez from ~/Movies/graduation-2026.mp4. Subtitle "Lincoln High School · Class of 2026".

Claude reads SKILL.md and drives the pipeline end to end. A 2-hour recording takes roughly 20–30 minutes total on an M1 (transcription ~10–15 min, music ~3 min, the rest seconds).

Prefer to drive it by hand? SKILL.md is also a human-readable runbook — every step is a plain command, and scripts/ contains the three helpers (transcribe_vad.py, make_title.py, gen_music.py), each with --help.

How it works (the short version)

  1. Find the moment — VAD-gated Whisper transcription (plain Whisper hallucinates "Thank you." over ceremony music; speech-only gating fixes it), then a surname-first search of the transcript.
  2. Cut a ~42-second window starting 5 seconds before the name — long enough to cover the name-to-walk lag — then extract frames and visually confirm the subject is the one crossing (one human confirmation beats any name-math).
  3. Title card rendered as a transparent PNG, composited with alpha fades.
  4. Score composed locally by MusicGen-small on the GPU.
  5. Mix with a deterministic volume envelope: music ducks to 18% under the name, swells for the walk, yields to natural applause.
  6. Verify — the name window of the final mix is re-transcribed; if Whisper can hear the name through the music, so can grandma.

Notes

  • Works for any ceremony with an announcer: graduations, recitals, award nights, citizenship ceremonies, weddings.
  • MusicGen's weights are CC-BY-NC — fine for personal/family clips, not for commercial use.
  • Whisper may spell the name creatively in transcripts ("Delilah" → "Dalila"); the skill searches tolerantly, and it doesn't affect the video.

License

MIT (the skill and scripts). Model weights retain their own licenses.

Ver o original no GitHub

Bom para

Relacionados