mimic-video-captionslisted

# Mimic On-Screen Captions The owner of "reproduce the captions of THIS video." A caption is its **motion + style on the footage**, not just its words — so a faithful reproduction needs THREE things, from TWO sources, then a **render-verify loop** to prove it: 1. **The exact words + timing** — authoritative from the **Whisper transcript** (`audio-analysis`). 2. **The visual treatment + how it animates** — from a **caption-focused paid analysis** (`extra_analysis_model({ focus: "captions" })`), which watches the WHOLE video so it can describe motion a per-frame still can't. 3. **The right renderer** for each caption — 3D/perspective vs flat-kinetic-2D vs plain subtitle. Then you **build → render → look at the pixels → fix** (the verify loop), because the agent builds blind otherwise and over-scales captions out of frame. > **Why a dedicated skill.** Reproducing captions is a craft with its own failure modes > (deciding 2D-vs-3D off frozen stills, mis-reading sung words, over-scaling 3D text out of > frame, building blind with no render feedback). `mimic-video` stays a thin router; this skill > owns the caption craft so the router stays clean. --- ## Step 1 — Split the two sources (words vs look). This split is the whole game. The video model **mis-reads words and timing** but **sees the visual treatment**. The transcript gives **exact words + word-level timing** but says nothing about look. So: - **Words + timing → the Whisper transcript.** Invoke `audio-analysis