← ClaudeAtlas

using-object-trackinglisted

Pin any visual element (emoji, text, image, video, JS draw fn, blur/pixelate/mask) to a moving subject across a video — face or object tracking, censor a face, follow a product. "smiley on her face", "blur the license plate", "logo on his shirt", "name tag follows him", "pixelate the plate".
Nagellabs/libi · ★ 2 · AI & Automation · score 75
Install: claude install-skill Nagellabs/libi
# Using Object Tracking > This skill is the ONE place for tracking + overlay-pinning. It absorbed the > old `tracking-overlay` skill — do not look for a separate overlay skill. > **Engine install:** If a tracking tool returns `tracking_engine_not_installed`, > follow the libi-tracking install plan: `libi.get_install_plan` > (`mcpId: "libi-tracking"`) → disclose ~10–20 min / ~2 GB and get the user's > approval → `libi.install_tracking_engine` (the actual installer) → > `libi.verify_install` until `ok:true` → retry. > **Do NOT gate on `chromium` / `mediapipe-vision`** — > those are the legacy MediaPipe path, not the default `yoloe+botsort` engine. The tracker is **shot-segmented and composable**. You drive it; it reports per-segment quality and *why* a segment failed. Never hand-guess pixel boxes. If the overlay is **static** (same rect the whole clip — not following anything), use `libi.add_overlay({ kind: "text" })` / `libi.add_overlay({ kind: "image" })` etc. instead; no tracking needed. ## Deliver what the user asked — never a workaround If the user asks for "a smiley **on her face**", the deliverable is a face-sized emoji **on the face for the whole clip**. When tracking is imperfect you have exactly two honest moves: 1. **Fix the track** (the repair loop below — re-anchor the bad window, switch method, skip a genuinely-impossible range), or 2. **Tell the user** what's still imperfect and ask how they want to proceed. You may **NOT** silently substitute a degrad