What problem does it solve? Producing B-roll for knowledge-style talking-head videos requires segmenting transcripts, choosing visual templates, writing generation prompts, and editing clips back onto the source video, all of which is slow and error-prone when done manually. ## Core Features & Use Cases - Transcript or Video Input: Accept a plain transcript or a talking-head video that is transcribed into a timestamped transcript via Whisper. - Semantic Routing and Templates: Split narration into semantic beats, route each to A-roll, real evidence, Hyperframes, or MiniMax-H3, and match one of 21 packaging templates (T01–T21) with three prompt profiles (v1 simple, v2 rich, v3 balanced). - Approval-Gated Generation: Produce broll-plan.json and broll-review.md, require explicit per-shot approval and a dry-run before any paid MiniMax-H3 API call, and forbid background music in all generated shots. - Automatic Editing: Insert approved B-roll back onto the source video by transcript timestamps while preserving the original speech track, exporting a final MP4 and edit-manifest.json. - Use Case: A creator uploads a 10-minute talking-head video, reviews a shot list of 8 proposed B-roll clips, approves 5 of them, and receives a final edited video with the generated clips inserted at the correct timestamps. ## Quick Start Use the minimax-broll-generator skill to analyze my talking-head video, plan B-roll shots with the v3 balanced profile, and after my approval generate and edit them into the final cut.
