// the find
rushindrasinha/youtube-shorts-pipeline
Automated YouTube Shorts pipeline: news → script → AI visuals → voiceover → captions → upload
Verticals v3 is a CLI pipeline that takes a topic string and outputs a published YouTube Short: research, script, b-roll, voiceover, captions, ffmpeg assembly, YouTube upload, ~$0.11 per video. It's aimed at solo creators or small teams who want to run a consistent short-form content operation without hiring an editor. This is not a creative tool; it's a production line.
The niche profile system is the genuinely interesting part — a YAML file that shapes tone, hook patterns, visual vocabulary, caption style, and music mood across every stage without prompt engineering per-run. Provider abstraction is well-thought-out: Claude, Gemini, GPT, Ollama, and MiniMax for LLM; five TTS options; all switchable per-run with a flag. Resumable stages via state.py mean a failed Whisper run doesn't throw away the Gemini images you already paid for. The research-as-hallucination-gate design (LLM instructed to use only scraped facts, not training knowledge) is the right instinct for a news-explainer pipeline.
The research layer is DuckDuckGo scraping with 300-char snippet truncation — for any fast-moving or nuanced topic, that's thin sourcing and the 'anti-hallucination gate' is weaker than it sounds since the LLM still has to infer from fragments. Visual quality is structurally limited: three Gemini Imagen frames with Ken Burns motion is functional but thin, and the roadmap items (Pexels stock footage, ComfyUI, Replicate) that would actually matter for visual variety aren't shipped yet. The `true_crime` in `politics` content niches, combined with automated upload defaulting to private-only-until-you-change-it, suggests the author knows this pipeline can produce misinformation at scale and has thought about it minimally. YouTube OAuth setup requires a local browser flow, which breaks any headless or CI use case without workarounds that aren't documented in the main README.