跳到正文

krakonjac300-pixel

podcast-shorts-factory

Ten cooperating AI agents that turn long podcasts into short-form videos, automatically. Free and open source, runs on free AI providers.

README 已保存到本站,可直接阅读

Documentation snapshot

README 快照

本页保存的是公开项目资料快照,阅读过程不需要连接 GitHub。

Podcast Shorts Factory

Ten cooperating AI agents that turn long podcasts into short-form videos, automatically.

Free and open source. It ships configured for free AI providers, so you can run the whole thing at roughly zero cost. No subscription, no paid API required.

You give it a YouTube podcast link. The agents transcribe the episode, find the moments most likely to perform, cut them to vertical 9:16, burn in animated captions, add music, b-roll and effects, quality-check every render, then write the titles and hashtags and schedule the posts to YouTube Shorts. It can run completely unattended on a daily schedule.


What each agent does

#AgentWhat it does
1FinderDownloads the podcast and transcribes it with word-level timestamps (faster-whisper), then uses AI to score the most clip-worthy moments. Long episodes are scored in chunks so nothing is missed.
2EditorCuts each clip with ffmpeg, reframes to vertical 9:16 with face tracking, burns in karaoke-style captions, and mixes music. Adds punch zooms on emphasis words, fake-multicam camera cuts, b-roll, a cinematic grade and a retention bar.
3UploaderWrites platform-specific titles, descriptions and hashtags, then uploads to YouTube Shorts through the official API. It can schedule a whole day server-side, so your PC can be off and the posts still go out. TikTok and Reels are exported ready to post.
4ManagerPulls your real metrics (views, retention), works out what is actually winning, and writes what it learns back into a file the Finder reads on the next run. This is the feedback loop that makes the system improve over time.
5Trend ScoutUses free web search to find what is trending in your niche right now, so clip selection leans toward what people are currently watching.
6PlannerGives each clip its creative direction: the on-screen hook text, music mood, which words to emphasise, and where to place sound effects, b-roll and transitions.
7CommunityDrafts replies to the comments on your posts, so engagement does not die on the vine.
8Finishing EditorThe quality gate. It watches every finished render the way a picky human editor would, and catches captions covering a speaker’s face, black or frozen frames, dead air, clipped or too-quiet audio, and wrong durations. It auto-fixes the cheap problems and blocks broken clips from ever being posted.
9TrainerThe team’s coach. Weekly, it studies the top-performing Shorts in your niche and updates one craft playbook, so the whole system levels up instead of standing still.
10CompilerThe showrunner. It stitches the week’s best moments into a long-form 16:9 episode with an AI-narrator editorial spine (thesis, per-clip analysis, verdict), designed to satisfy 2026 reused-content monetization rules.

On top of the agents there are 24 skill playbooks in factory/skills/: written craft knowledge on hooks, storytelling, pacing, sound design, captions, thumbnails, SEO, engagement and monetization. These get loaded directly into the agents’ prompts, which is what makes the output feel deliberate instead of random. Edit them freely, they are just markdown.


Quick start

Prerequisites: Python 3.11+ and ffmpeg.

winget install -e --id Python.Python.3.11
winget install -e --id Gyan.FFmpeg

Install:

  • Windows: double-click install.bat
  • macOS / Linux: bash install.sh

The installer builds the environment and then launches an interactive setup wizard that asks you a handful of questions: which AI provider to use, your API key, your niche (pick from presets), which platforms to post to, and your call-to-action. It writes the config for you, so there are no files to edit by hand.

Then run it:

.\factory auto "https://www.youtube.com/watch?v=VIDEO_ID"
./factory.sh auto "https://www.youtube.com/watch?v=VIDEO_ID"

Or go fully hands-free: set your source channels in config.yaml under scheduler.sources, then run .\factory daily.

A full step-by-step walkthrough is in SETUP-GUIDE.pdf, and QUICKSTART.md has the same thing in text form.


It runs free

The default configuration points at free AI providers. Pick one, paste the key in during setup, and you are running:

ProviderCostGet a key
OpenRouterfree tierhttps://openrouter.ai/keys
Groqfreehttps://console.groq.com/keys
Google Geminifreehttps://aistudio.google.com/apikey
Ollamafree, localhttps://ollama.com
Anthropic (Claude)paid, best qualityhttps://console.anthropic.com

Transcription runs locally with faster-whisper, the voiceover uses free Edge TTS, and the background music library is generated with pure synthesis, so it is safe to monetize.


Commands

setup          First-run wizard (provider, key, niche, platforms, CTA)
find      Download, transcribe, and AI-score clip candidates
review         Approve or reject candidates yourself
edit           Render approved clips into vertical shorts
finish         QA-review and finish each render before it can post
upload         Post or export to platforms
compile        Build the weekly long-form episode
stats          Refresh metrics and update learnings
scout          Refresh current trends
skills         List installed skills and which agent uses each
auto      The whole pipeline in one command
daily          Fully unattended: trends, newest video, clips, edit, post

Why I am giving this away

I built this over almost a month. It started because I wanted a channel that could run itself, and it turned into a proper little team of agents that argue with each other about what makes a clip good.

I am releasing it free because I would rather it be used than sit on my drive. If one of you takes this and actually manages to monetize it, that genuinely makes me happy. That is the whole hope. The MIT license means you can use it commercially, modify it, and sell whatever you build with it. You do not owe me anything.

If you do build something with it, I would love to hear about it.


Contributing

Issues and pull requests are very welcome. Useful places to start:

  • Better clip-selection prompts in config.yaml and factory/skills/
  • More editor styles and transitions
  • Platform support beyond YouTube (direct TikTok and Instagram posting)
  • Anything in the Finishing Editor’s quality checks

Please use it responsibly

This automates production, not judgement. Only clip content you have the rights to: your own material, licensed material, fair-use commentary, or content you have permission to use. You are responsible for the platform rules you operate under and any API costs you incur. Fully automated posting can attract strikes if you are careless, so start with privacy: private, watch what it produces, and only then go public.


Disclaimer

This is not perfect. It is a working system that produces real videos, but it has rough edges, the output quality varies, and some parts are held together with duct tape and stubbornness. I am releasing it as-is, honestly.

Hopefully, with the help of the community, we can make it perfect for everyone. Every fix, idea and pull request genuinely helps.


License

MIT. See LICENSE. Use it, change it, sell it, have fun.

Official distribution

获取与安装

暂未发现可确认的官方软件包地址

当前 README 快照没有出现 npm、PyPI、Crates.io、pub.dev 等官方包页链接。本站不会根据仓库名称猜测下载地址。

本站不托管项目文件;需要安装时,请以项目维护者发布的官方文档为准。

使用前核验

本站保存公开资料用于阅读,不代表安全审计或功能背书。安装前请核对许可证、依赖来源和发布签名,不要直接运行来源不明的二进制文件或高权限脚本。