Claude 与 GPT-6 Astra 接 DaVinci Resolve MCP 实测对比
Claude 适合单次强指令剪辑,Astra 配 Gemini 视觉可循环自审,1 小时 55 分完成 4 条成片且 token 仅占 3%。
ran both Claude and GPT-6 Astra against DaVinci Resolve's native MCP server, here's what actually separates them Claude + DaVinci: - built a full project from scratch, imported media, synced dual-system audio, de-camed a film scan into individual shots - assembled a 15-minute rough edit from raw footage, correctly reordering scenes to follow a process - automated a 10-20 hour TIFF cache render + final H.265 export unattended, recovering from a Resolve crash mid-run - generated 20 vertical cut-downs from a single horizontal timeline with smart reframe applied automatically, one unresolved gap issue due to an API limitation, not Claude's error - workflow here is single-shot, heavily instructed step by step GPT-6 Astra + DaVinci: - paired with Gemini for actual visual scene understanding, not just transcription, so it can identify what's literally in frame (a whiteboard, a room, people) and auto-censor sensitive content it sees on screen - run as a loop instead of a single instruction: given one large prompt with all source files, tools, and editing preferences, then left to work, review its own output, and revise autonomously until done - produced 4 fully edited training videos end to end, b-roll, music, image generation, captions, in 1 hour 55 minutes - token cost was dramatically lower in this workflow, 3% of weekly usage vs a much heavier spend on the code-based editing method used previously the real fork isn't which agent is smarter, it's workflow shape Claude's DaVinci demo is a single, tightly scoped session, best when you already know exactly what you want and want to direct it step by step Astra's DaVinci demo is a loop with Gemini bolted on for vision, best when you want to hand off a full batch of raw footage and voiceovers and come back later to a finished draft both are still assistant editors, not replacements, procedural work (syncing, organizing, rough cuts, rendering) is what's actually being automated here, the creative judgment on what makes a good cut still needs a human pass
这条是引用着另一条写的:x.com/rewind02/status/2100536800204693623