Making a video podcast is hard not because of editing software, but because of the dozens of decisions between topic and finished video: research, scripting, tone, visuals, subtitles, BGM, thumbnails, titles, tags. video-podcast-maker by Agents365-ai packs the entire chain into a single SKILL.md — Claude Code acts as director, Remotion is the studio, and seven TTS engines are the voice actors. You say "make me a video about X" and the agent handles the rest.
Key points
- One command, topic to finished video: the pipeline runs 14 steps — topic definition → web research → script writing → pronunciation correction → TTS narration → timeline sync → Remotion render → BGM overlay → subtitle burn-in → thumbnail generation → multi-platform publish info. Each step's output feeds the next.
- Project structure: each video gets one directory containing topic notes, research, narration script (
podcast.txt),phonemes.json(polyphone dictionary), TTS audio, subtitles (podcast_audio.srt),timing.json, thumbnails, per-platformpublish_info.md, andfinal_video.mp4(native 4K, 3840×2160). - 7 TTS backends: Edge TTS (free), Azure Speech, Volcano Engine Doubao, Alibaba CosyVoice, ElevenLabs, Google Cloud TTS, OpenAI TTS — switched via
TTS_BACKENDenv var. Azure supports Chinese-English mixed narration natively; a global plus per-projectphonemes.jsonfixes Chinese polyphones. - Remotion rendering: React components define layouts, CSS defines animation; reusable templates (ComparisonCard, Timeline, CodeBlock, QuoteBlock, FeatureGrid, DataBar, StatCounter, FlowChart, IconCard). Real-time preview via
npx remotion studio.[SECTION:title]markers auto-generate chapter progress bars. - Platform differentiation, not cropping: Bilibili (16:9, dense visuals, timestamps, dual 16:9+4:3 covers), YouTube (SEO titles, chapter links, subscribe CTA), Xiaohongshu (3:4, emoji-friendly 200-500 char posts, double-hashtag topics), Douyin (9:16 vertical highlights only, 1-minute cuts), WeChat Channels (9:16, share-oriented CTAs). Same topic, five rewritten versions.
- Human final review of scripts: the SKILL.md explicitly requires users to read the AI draft aloud, revise at least three times (typos → cut fluff → pacing), and separately audit numbers, proper nouns, and English terms — where 90% of TTS mispronunciations occur.
- Preference learning: color/font/speed choices in Remotion Studio are saved to
user_prefs.jsonand auto-applied in future projects. Roadmap includes design-reference libraries, Playwright-based visual-style scraping, and style-profile matching. - Remotion customization requires some React/CSS knowledge for deep visual tuning.
- TTS still lacks human emotional delivery — real narration remains irreplaceable for personal-brand podcasts.
- AI thumbnails can look "AI-generated" and need manual curation.
- Multi-platform publishing is semi-automatic: files and metadata are generated, but uploading is manual (YouTube Data API automation on roadmap).
- Verified on macOS; Linux compatible; Windows/WSL support unconfirmed.
- GitHub: https://github.com/Agents365-ai/video-podcast-maker
- Chinese README: https://github.com/Agents365-ai/video-podcast-maker/blob/main/README_CN.md
- Skill marketplace: https://github.com/Agents365-ai/365-skills
- Author: Agents365-ai — Bilibili: https://space.bilibili.com/441831884
- Discord: https://discord.gg/79JF5Atuk
- Supported agents: Claude Code, Codex, OpenCode, OpenClaw (ClawHub)
- License: MIT
Architecture
| Layer | Tool | Rationale | |---|---|---| | Orchestration | SKILL.md + Claude Code | Natural-language driven, no coding | | Rendering | Remotion (React) | Code-as-animation, reusable components, native 4K | | Audio | TTS engines + FFmpeg | Multi-backend, BGM, subtitle burn-in | | Thumbnails | AI (Imagen/Qwen) / Remotion | 16:9 + 4:3 versions | | Dev | Remotion Studio | Live visual editing | | Deploy | Docker / local Node | No cloud lock-in |
Dependencies: Python 3.8+, Node.js 18+, FFmpeg 4.0+. Zero-cost startup via free Edge TTS.
Limitations
Why it matters
A 10-minute video that once took 15+ hours (research, scripting, recording, editing, thumbnail, description) can now take 2-3 hours: one-sentence topic → agent research → AI script → 30 minutes of human review → TTS → Remotion render → auto cover and publish info. Human effort shifts from "how to edit" to "what to say" — individual creator productivity rises 5-10x, and competition shifts from diligence to insight.