A Claude Code skill that turns one topic prompt into a finished Vox-style explainer video. Script, keyframes, animation, voiceover, music, and local assembly — end to end.
One topic in, finished film out.
TOPIC → Script → Voiceover → Keyframes → Animation → Music → Assembly
(Claude) (xAI TTS) (Seedream (Gemini (MiniMax (ffmpeg)
5.0 Pro) Omni Flash) Music 2.6)
All generation runs through the Atlas Cloud unified API — one key covers every model. Assembly happens locally with ffmpeg.
# Claude Code
mkdir -p ~/.claude/skills
git clone https://github.com/YOURUSER/vox-explainer ~/.claude/skills/vox-explainerRequirements:
ATLASCLOUD_API_KEYin your environment (free credits on signup)ffmpeg+ffprobewith libass- Python 3.10+ with
requests
Open Claude Code and ask:
Make me a 60-second Vox-style explainer about the Tang Dynasty golden age.
Claude reads the skill, writes the script in beats, generates and measures narration first (VO duration drives all timing), locks a visual style with a Seedream anchor frame, animates each keyframe with subtle collage motion, scores it, and assembles a subtitled, mixed, faststart MP4.
| File | What it covers |
|---|---|
SKILL.md |
The six-stage pipeline, project structure, orchestration rules |
references/vox-style-guide.md |
The writing arc and the mixed-media collage visual grammar, with prompt templates |
references/atlas-cloud-api.md |
Endpoints, model IDs, async polling pattern, cost budgeting |
references/ffmpeg-assembly.md |
Conform, concat, ASS subtitles, sidechain music ducking, final encode |
- VO-first timing. Narration is generated and measured before any video exists. Beat durations become the master clock, so clips are conformed to speech instead of speech being squeezed into clips.
- Style anchor pattern. One approved anchor frame + Seedream's edit endpoint (reference images) keeps palette, texture, and framing consistent across every keyframe — the thing that makes it look art-directed instead of generated.
- Subtle motion vocabulary. Explainer collage should move like motion graphics, not footage. The skill encodes an allowed/forbidden motion table and forces negative constraints into every animation prompt.
- Everything is resumable. All state lives in project files (
script.json,vo_durations.json, per-beat assets), so any single beat can be tweaked and regenerated without touching the rest.
Roughly $3–8 per 60-second film at July 2026 Atlas Cloud pricing; animation dominates. The skill quotes a budget before the first expensive stage and offers a cheap-tier draft pass.
MIT