Nothing screams automated slop faster than bright white captions parked directly across a founder’s nose, or a call-to-action badge tucked right behind Instagram’s heart and share icons.
That mistake happens every single day because standard FFmpeg drawtext filters and basic local video scripts are completely blind to what is actually inside the frame. They hardcode y = height * 0.78, render the MP4, and pray the presenter stood in the exact center of the room.
Teams building local coding agents try to patch that blindness by adding a self-critique loop: render the video, extract twenty PNG screenshots, upload all twenty images into Claude Opus 5.5, Fable 5.2, or GPT-6.2 Astra, ask the model whether any text overlaps a face, edit the CSS, and repeat.
That vision loop works eventually, right up until you check your billing console and see 180,000 multimodal tokens burned just playing twenty questions with static screenshots at $8/$40 frontier rates.
Visual quality assurance belongs right next to the render worker, not inside your chat client’s token meter.
When your agent calls generate_preview (or revise_preview after an edit note) on https://mixi.mixbash.com/api/agent/mcp, Mixbash executes motion authoring and immediately runs an independent, server-side frame-level visual QA pass before marking the preview ready.
That worker inspects rendered frames against platform safe-zone maps (keeping text clear of Reels captions, Shorts subscribe buttons, and top status bars), checks subject and face positioning, verifies text bounding boxes for clipped characters or contrast drop-offs, and resolves layout collisions automatically.
Your local model, even a lightweight runner like GPT-6 Luna or Claude Haiku 5.5, simply receives a compact JSON confirmation and fetches the clean browser link via get_project_page. Zero PNG screenshots ever enter your conversation context.
Still, let’s keep expectations grounded about what automated visual QA can and cannot judge. Server-side QA is ruthless about geometric mistakes, clipped descenders, safe-zone trespassing, and face-caption collisions, yet it will not tell you that your presenter looked slightly sarcastic while thanking customers at 0:14.
That human nuance is why get_project_page gives you an instant half-resolution browser preview to watch with your own eyes before you ever authorize approve_review and render_final.