In this video, the creator shares their workflow for reverse engineering AI-powered video clipping tools, using Opus Clip as an example. They begin by explaining the initial steps, which involve downloading a YouTube video using tools like yt-dlp. The video is then processed by extracting the audio, which is transcribed into text with timestamps. This transcription serves as the foundation for identifying interesting moments within the video that can be clipped.
Next, the creator discusses leveraging large language models (LLMs), specifically Gemini 3, to analyze the transcription. Gemini 3 helps generate ideas for potential clips by identifying engaging or relevant segments in the transcript. These segments are then organized into a timeline JSON file, which includes start and end timestamps, clip text, and tags such as “funny” or “educational.” This structured data guides the subsequent video processing steps.
The video processing involves cutting the original video into smaller clips based on the timeline JSON using ffmpeg. The creator also highlights the importance of converting horizontal videos into vertical formats, such as 9:16, which are popular on social media platforms. To achieve this, they suggest using YOLO, an object detection model, to detect faces in the clips and crop the video around the speaker’s face. This ensures the focus remains on the person talking, enhancing viewer engagement.
Additionally, the creator mentions adding captions to the clips by burning subtitles onto the video using ffmpeg. This step improves accessibility and viewer comprehension. They then demonstrate how they use Claude Code, an AI coding assistant, to generate a detailed plan for the entire pipeline, from downloading the video to producing the final clipped and cropped output. The plan includes manual clip selection, local transcription with Whisper, speaker detection, and face cropping.
Finally, the creator runs the pipeline on a sample Joe Rogan video, showing the process from downloading to transcription, clip suggestion, face detection, and cropping. While the result is not as polished as Opus Clip, it successfully demonstrates the core workflow. The video concludes with reflections on the iterative nature of reverse engineering and encourages viewers to experiment with AI tools to build their own video clipping pipelines. The creator emphasizes that this approach can be refined further but serves as a practical starting point for similar projects.
