video-talkcraft

video-talkcraft is a voiceover-video Skill by Vincent Wei (GitHub user Vincentwei1021), part of his video-shotcraft series; the README says it turns Claude Code or Codex into a motion studio for talking-head and narrated videos. You supply a script and a finished voiceover (a real recording or any TTS), and it aligns the script to the audio on your own machine down to each word's timing, writes every semantic beat into a storyboard called a SHOTBOOK, then picks from more than a hundred motion recipe cards, such as kinetic type, data shots, web-page evidence, camera moves and transitions, and renders the video with Remotion, with captions and sound effects locked to the voice. Real recordings can first have filler sounds and stumbled restarts trimmed automatically, and after delivery it opens a local workbench, similar to the CapCut editor, where you can fine-tune each element on a multi-track timeline. The default is landscape 1920×1080, switching to 1080×1920 for vertical platforms such as Douyin; it's designed for Chinese narration and handles mixed Chinese and English. What sets it apart is how hard it works to avoid a "talking over slides" feel: every motion beat is anchored to an exact word in the voiceover, scripts flag any frozen stretches after rendering, and one independent review pass happens before delivery.

Who it's for:creators making science explainers, course videos, research walkthroughs or personal commentary videos for noncommercial use only. You'll need Claude Code or Codex and a willingness to install a large set of local dependencies.

Before you start: the license is PolyForm Noncommercial 1.0.0, so it is for noncommercial use only. Personal, educational and research use is free, and any commercial use of the tool requires the author's permission in advance, which the README says you can request by email or GitHub issue; the LICENSE states that videos made with it belong to their creators. Remotion, which does the rendering, has its own license terms to review separately. Install with npx skills add Vincentwei1021/video-talkcraft; you need Node 18 or later, Python 3.10 or later and ffmpeg. The downloads are sizeable: the dependency check that runs at the start of each new video installs or upgrades Remotion and a headless browser over the internet, about 760 MB of shared dependencies including roughly 190 MB for the browser, with every video project symlinked to that one copy. The default local alignment model, FireRedASR, is a one-time download of about 767 MB, or faster-whisper downloads about 460 MB automatically on first run. By default it uses your own finished audio and aligns it locally; if you have no recording, you can opt into Fish Audio for synthesis with your own API key, which sends the script and voice ID to Fish Audio, with free quota and pricing set by Fish Audio. When gathering footage it downloads B-roll video and images from sites such as Pexels and Pixabay, whose APIs need keys you request yourself, and it uses Playwright to open and capture real web pages. The footage guide only lists sources that don't require attribution, but you still need to check each asset's license, since the tool's license doesn't cover the footage. It doesn't generate digital humans; any on-screen presenter comes from footage you provide. The presenter in the repo's demos is an AI-generated placeholder, and the README says to replace it with your own footage for real productions; get consent before using anyone else's likeness or voice.

Based on the project's public documentation. Not tested by AI Inspiration World.