take-geometrylisted
Install: claude install-skill vaibhav-barman/live-clipper
# TAKE GEOMETRY (stream edition) — measure, then solve the layout
> **THE JOB:** before a single graphic is placed, know where the face, hair and hands actually are
> across the WHOLE clip — as distributions, not as one frame — and derive the card fit and the clear
> text bands arithmetically. Every layout rejection in the original project traced back to a number
> that was guessed or sampled too thinly.
**Nothing transfers between clips.** A streamer leans in, a scene switches, the facecam moves between
"starting soon", gameplay and "just chatting" scenes. Measure every clip. Cache per **stream + scene**
only after verifying it still holds.
Adapted from the creator's skill. **Kept:** distributions over single frames, the detector warnings,
solve-don't-nudge, headroom, the alpha guard, writing the numbers down. **Added:** Step 0 (find the
facecam inside a composited stream frame) and the facecam upscale limit.
## The two hard rules this serves
1. **Text lives inside the safe zone**: below y=220, above y=1470, 35 px sides, ~100 px clear on the
right for the action rail. Imagery may cross; TEXT may not.
2. **Text must NEVER cover a face in ANY frame.** A ~6-frame flash is a defect. When no face-clear
spot exists, **drop the element** rather than overlap it.
## Step 0 — find the facecam (stream only)
The facecam is a rectangle composited into the 16:9 stream frame, usually in a corner, often with a
border or rounded mask.
1. Sample ~20 frames across the clip. Run a