What I had to write down before an LLM video editor stopped looking automated

A developer found that directly prompting an LLM to edit videos with ffmpeg resulted in machine-made looking content, with persistent issues like repeated words at seams and inverted sentences from mid-clause cuts. To overcome this, they stopped feeding the video directly to the model. Instead, they used whisper.cpp for local transcription with word timings and ffmpeg silencedetect for boundary proposals. The LLM then processes the text transcript to generate a cut list, with audio serving as the primary clock. Playwright is used to capture content from URLs mentioned in the script, ensuring accuracy.