Video Maker

English | 简体中文

video-maker is a Codex skill for end-to-end Bilibili explainer-video production on Windows.

Repo map

• SKILL.md: entrypoint and operating rules
• scripts/bootstrap_project.py: full new-project scaffold plus runtime helper generation
• scripts/bootstrap_video_project.py: low-level content scaffold used by the full bootstrap wrapper
• scripts/upgrade_project.py: upgrade existing projects and rewrite runtime helpers
• references/agent-orchestration.md: simplified content-strategist -> script-writer -> narration-polisher content pipeline
• references/bilibili-tech-explainer-workflow.md: beat-driven explainer workflow
• references/narration-polisher.md: 口播润色 subagent 的职责和推荐提示词
• remotion-best-practices skill: default engine for video composition, animation, subtitles, audio mounting, timeline, and rendering
• web-design-engineer skill: optional visual-design support for layout and motion taste; production output stays in Remotion
• references/quiet-glass-lab-v3.md: iOS 18-inspired frosted glass brand prompt pack
• references/quiet-glass-lab/base.css: neutral render foundation only, not a visual theme or module template
• references/chinese-voice-rules.md: narration consistency rules
• references/video-acceptance-rubric.md: final QA gates including meaning-gain checks
• publish/cover_prompt.md in generated projects: imagegen-ready Bilibili cover prompt template
• content/visual_qa_report.json in generated projects: Remotion frame-sample and cover visual repair report
• references/imagegen-2-visual-playbook.md: GPT Image 2 / imagegen production rules for direct-rendered visuals with visual QA and regeneration
• scripts/doctor.py: environment checker

Model

Default flow:

1. coordinator creates the project, runs simple commands, assigns subagent task packets, and makes final go / no-go decisions.
2. content-strategist writes problem, audience, opening, meaning, outline, depth, detail, and evidence contracts.
3. script-writer writes content/script_draft.json.
4. narration-polisher polishes the finished draft into natural Bilibili narration in content/narration_polish.json.
5. coordinator compiles shot intents and render-ready segments from the approved narration.
6. visual-architect designs Remotion scenes and uses imagegen for visual assets, exploded diagrams, animation assets, and the cover base.
7. visual-qa-fixer inspects Remotion frame samples and publish/cover.png, then fixes Remotion code/assets/cover text or regenerates imagegen outputs until visual blockers are gone.
8. production-engineer owns Qwen master-track audio, Remotion props, timeline, render, and export.
9. acceptance-reviewer checks many real screenshots, reads key Remotion code and content contracts, and listens to required audio samples.
10. Publish handoff runs only after video, cover, metadata, visual QA, and acceptance all pass.

Stable defaults:

• new projects do not use HTML slide rendering; Remotion scene code and props appear after content and shot intents are locked
• the main agent is a coordinator, not a maker; substantial content, visuals, voice, assembly, and review are delegated
• video composition, animation, subtitles, audio mounting, timeline, and rendering default to Remotion; the visual system is a brand prompt pack, not fixed scene templates
• visual assets and Bilibili covers default to visual-architect + imagegen, with final rendered-pixel repair owned by visual-qa-fixer
• visual production records benchmark and key visual inside visual_asset_plan so inspiration, cover promise, reusable key visual, imagegen assets, and QA stay connected without extra contract files
• Chinese narration prefers local Qwen first
• narration uses a single master-track path with a hard timeout and voice_jobs/qwen_master_status.json status manifest
• full-audio approval requires voice_profile.full_audio_review_status = passed; opening or midpoint spot checks are only supporting notes
• Chinese pace target is calibrated to the 2026-04-25 CUDA video: about 260 CJK chars/min, with duration mismatch treated as non-blocking when content and QA pass
• content planning now starts from problem_contract, then writes evidence-backed contracts before any render-ready segment exists
• finished content drafts are always sent through narration-polisher to remove AI machine-tone and lower ordinary viewer comprehension cost
• Bilibili cover generation happens after title lock and local QA, not during early content phases
• prepare_publish_job.py now emits cover_path and refuses publish handoff when the cover file is missing

For real Bilibili publishing, pair this skill with desktop-control-for-windows.

See SKILL.md for usage, quiet-glass-lab-v3.md for brand visual rules, and chinese-voice-rules.md for voice rules. Remotion execution follows the companion remotion-best-practices skill.

License

MIT