convez.ai
SkillAvançado
GiomarDev

ReelForge — Studio de Vídeo & Legendas com IA

GiomarDev/ReelForge

Studio completo de vídeo: legendas síncronas, tradução, motion graphics e B-roll automático

Ver no GitHub
InicialPluginScore · de 0 a 100

Skill para Claude Code que transforma clipes brutos em vídeos profissionais com legendas sincronizadas em nível de palavra (estilo TikTok/Reels), tradução automática, gráficos animados e efeitos visuais. Integra HyperFrames, Whisper e FFmpeg para renderizar MP4 final localmente, sem dependências na nuvem.

Instalação

$npx convez-ai GiomarDev/ReelForge

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
0
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 5 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Criar conteúdo viral para redes sociais (TikTok, Reels, YouTube Shorts) com legendas profissionais
  • Traduzir e legendar vídeos em múltiplos idiomas mantendo sincronização perfeita
  • Automatizar produção de vídeos educacionais com gráficos animados e overlays

Veredito

ReelForge é uma ferramenta robusta e determinística para produção de vídeo em larga escala dentro de Claude Code. Excelente para criadores que precisam de legenda sincronizada e motion graphics, mas requer conhecimento técnico sólido (Node.js, ffmpeg, CLI) e infraestrutura local. A qualidade de saída é profissional e a automatização é genuína.

README

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

🎬 ReelForge

An AI video & caption studio for Claude Code. Hand Claude a raw clip and it installs the toolchain, transcribes the speech, burns in word-synced (TikTok/Reels-style) captions, translates them into any language, adds motion graphics — charts, maps, social cards, shader transitions, AI voiceover — and renders a finished MP4. All locally, all deterministic.

Powered by HyperFrames (HeyGen's HTML-to-video renderer), Whisper, and FFmpeg.

ReelForge ships as a Claude Code skill: drop it in, and Claude knows the entire pipeline — the commands, the gotchas, and a battle-tested caption generator — out of the box.


✨ What you can do

  • 🗣️ Auto-subtitles — word-level captions synced to speech, with a moving highlight.
  • 🌍 Translate captions — keep the original audio, show subtitles in another language.
  • 📈 Motion graphics — animated bar/line charts, choropleth maps, flowcharts.
  • 🎞️ Real B-roll cutaways — cut to royalty-free stock footage (Pexels) on the words the speaker says, with trendy transitions (zoom-punch, whip-pan, flash, glitch) — audio never breaks.
  • 📱 Social overlays — TikTok/Instagram/YouTube follow cards, X & Reddit posts, Spotify cards.
  • 🎞️ Transitions & atmosphere — shader wipes, film grain, vignette, logo outros.
  • 🔊 AI voiceover — local text-to-speech with Kokoro-82M.
  • ✂️ Background removal — matte out a subject for transparent compositing.
  • 🌐 Website-to-video — capture a live page as footage.
  • 📤 Render anywhere — MP4, transparent WebM/MOV overlays, PNG sequences, or HeyGen cloud.
  • ✅ Virality QA loop — the viral-check skill audits the result (format, hook, captions, pacing, loudness, safe areas, trend fit) and iterates fixes until there are no blockers.

🚀 Quick start

1. Install the skill

# Copy the skill into your Claude Code skills directory
cp -r hyperframes-video ~/.claude/skills/

Then just talk to Claude: "add Spanish subtitles to the video in my Downloads" — it takes it from there. Or run the pipeline yourself with the commands below.

2. Install the toolchain (one time)

nvm install 22 && nvm use 22          # HyperFrames needs Node 22+
brew install ffmpeg whisper-cpp        # ffmpeg + whisper-cli
npx hyperframes doctor                 # verify everything is green

3. Caption a clip end-to-end

DUR=$(ffprobe -v error -show_entries format=duration -of csv=p=0 clip.mov)

# scaffold a project around the video
npx hyperframes init subs --video clip.mov --resolution portrait-4k --non-interactive --skip-skills

# transcribe to word-level timestamps (large-v3 = accurate, multilingual)
hyperframes-video/scripts/transcribe.sh clip.mov --lang en --model large-v3 --out subs/build/words.json

# generate the synced caption composition (TikTok style)
node hyperframes-video/scripts/generate-captions.mjs \
  --words subs/build/words.json --video clip.mov \
  --out subs/index.html --resolution portrait-4k --duration "$DUR" --download-font

# verify, then render
cd subs
npx hyperframes lint
npx hyperframes snapshot --at 12,24          # PNG previews — eyeball before rendering
npx hyperframes render -o ~/Downloads/clip-subtitled.mp4 -f 30

Translate the captions

Run the generator once, copy the printed line list into a JSON map, then re-run:

// translation.json  — map each English line to its translation
{ "Every great video starts": "Cada gran video empieza",
  "with a single frame.":     "con un solo fotograma." }
node hyperframes-video/scripts/generate-captions.mjs \
  --words subs/build/words.json --video clip.mov --out subs/index.html \
  --resolution portrait-4k --duration "$DUR" --translate translation.json --download-font

🧰 Command reference

CommandWhat it does
npx hyperframes doctorcheck Node / FFmpeg / Chrome
npx hyperframes init NAME --video IN.mov --resolution portrait-4kscaffold a project around a clip
scripts/transcribe.sh IN.mov --lang es --model large-v3 --out words.jsonspeech → word-level JSON
scripts/generate-captions.mjs --words words.json --video IN.mov --out index.html ...build synced caption composition
PEXELS_KEY=… scripts/fetch-broll.mjs --spec broll.json --out assets/brolldownload royalty-free B-roll clips for cutaways
npx hyperframes catalog / add <name>browse / install blocks & components
npx hyperframes tts "text" -o vo.wavlocal AI voiceover (Kokoro-82M)
npx hyperframes remove-background INmatte out background → transparent media
npx hyperframes capture URLrecord a website as footage
npx hyperframes lint / preview / snapshot --at 5,15validate / live studio / key-frames
npx hyperframes render -o out.mp4 -f 30 -q highrender MP4 (or --format webm/mov/png-sequence)

generate-captions.mjs flags: --resolution · --duration · --fps · --max-words · --gap · --font-size · --bottom · --highlight · --translate · --download-font. Full docs in hyperframes-video/SKILL.md and the references/ folder (catalog, captions deep-dive, composition rules, troubleshooting).

📦 What's in the box

Two Claude Code skills:

ReelForge/
├── hyperframes-video/            ← BUILD: raw clip → captioned, motion-graphic video
│   ├── SKILL.md                  ← the end-to-end workflow Claude reads
│   ├── scripts/
│   │   ├── transcribe.sh         ← audio → whisper.cpp word-level JSON
│   │   ├── generate-captions.mjs ← word JSON → synced caption composition
│   │   └── fetch-broll.mjs       ← Pexels → trimmed local B-roll clips
│   └── references/
│       ├── catalog.md            ← every block/component + media command
│       ├── captions.md           ← timing repair, styling, translation, fonts
│       ├── composition-rules.md  ← data-attributes, timelines, the 5 rules that bite
│       ├── viral-edit.md         ← hook + concept SVGs + real B-roll cutaways + transitions
│       └── troubleshooting.md    ← real failure modes & fixes (HDR/--sdr, rotation, …)
└── viral-check/                  ← REVIEW: audit a clip for virality, iterate until it passes
    ├── SKILL.md                  ← the see → critique → fix → re-render loop
    ├── scripts/
    │   └── analyze.mjs           ← objective scan: format, LUFS, pacing, hook accent
    └── references/
        └── rubric.md             ← the virality rubric (LatAm/Peru-aware)

hyperframes-video makes the video; viral-check grades it against a virality rubric and loops fixes until there are no blockers (vertical, hook, captions, pacing, sound, safe areas, ending, trend fit) — then masters loudness to ~-14 LUFS.

📋 Requirements

macOS · Node 22+ · FFmpeg · whisper.cpp · Google Chrome (used by HyperFrames to render).

📄 License

MIT — see LICENSE.


Not affiliated with HeyGen or Anthropic. HyperFrames, Whisper, and Claude are trademarks of their respective owners.

Ver o original no GitHub

Bom para

Relacionados