I Built an AI-Powered YouTube Shorts Generator (vitaliihonchar.com)

A developer released youtube-shorts-creator, an open-source CLI that automatically converts long YouTube videos into ready-to-publish Shorts. The pipeline extracts audio with ffmpeg, runs a local Whisper model to produce transcripts with per-word timestamps, sends that JSON to an LLM (GPT-4‑mini) to identify high-impact time ranges and generate metadata (title, description, tags), then cuts and applies effects with ffmpeg and uploads batches to YouTube. Example CLI: 1shorts-creator -v "video.mp4" -s 5 -sd 60 --upload produces five 60s shorts and uploads them automatically. For the AI/ML community this is a practical demo of LLM+multimodal orchestration for content repurposing: local speech models cut costs/privacy risk, while a small LLM guides semantic selection and metadata creation. Key technical choices—ffmpeg for performant editing vs. slow Python-native libraries, Whisper for timestamped transcripts, Strategy-pattern pluggable video effects (audio normalization, vertical ratio conversion, captions, blur, speedup), and LLM prompts for segmenting—make the system production-ready and extensible. The project highlights trade-offs (transcription errors needing LLM cleanup, prompt engineering, SEO tuning) and suggests the same architecture can be ported to Go for lower-runtime complexity. Code and prompts are available on GitHub for teams wanting to automate short-form video pipelines.