What problem does it solve? Building a complete video (voiceover, word-level captions, 9:16 reframing, mixing, muxing) requires coordinating many separate tools, and it is unclear which components belong in the browser extension versus a local sidecar. This Skill defines the architecture, the script.json contract, and the tool choices so the heavy video work runs locally outside the extension. ## Core Features & Use Cases - script.json contract: A single source of truth defining version, renderer, aspect, voice settings, and 3-12 scenes with voiceText, templateId, and inputs. - Sidecar tool selection: Maps each pipeline stage to a tool — audio.cpp for TTS and mixing, whisperX for ASR with word-level timestamps, PySceneDetect/MediaPipe patterns for 9:16 reframe, ffmpeg for concat and mux. - Flow MCP integration: Pulls images and b-roll from the SEOSONA Flow extension through a local MCP server with a validated FlowAsset/SceneInput contract, degrading gracefully when MCP is absent. - Use Case: When planning a Vietnamese short-form video product, use this Skill to define the scene contract, choose offline TTS/ASR binaries, and wire per-scene idempotent regeneration before building the sidecar CLI. ## Quick Start Ask the assistant to design the Video AI V2 pipeline for a 9:16 video with Vietnamese TTS, word-level captions, and b-roll sourced from the Flow MCP server.
