This repository is an agent skill: a self-contained workflow that turns one topic into a finished Vox-style paper-collage video (script → collage keyframes → motion → voice-over → music → captions). It is not tied to any single assistant — any coding agent that can read instructions and run scripts can drive it.
- Read
SKILL.md— the full workflow and the two human approval gates. (SKILL.zh.mdis the same in Chinese.) - Before writing any prompt, read
references/(prompt structures, the vocabulary/theme bank, and the narrative-beat library). - Work one project at a time under
out/<project>/, driven by a singlebeats.json. Run the stages inscripts/in order:style_bakeoff.py → keyframes.py → clips.py → audio.py → assemble.py.
ATLASCLOUD_API_KEYin the environment — https://www.atlascloud.ai/console/api-keysffmpeg+ffprobe- Python 3 with
pillow
- Claude Code auto-loads this as a skill from
SKILL.md's frontmatter — just ask for a "vox video". - Codex / other agents: follow
SKILL.mdas your instructions; thisAGENTS.mdis your entry point.