How do I install the Switch Faces skill?

Run npx skillmds@latest add jperrello/switch-faces in your terminal (requires Node.js), paste this page's agent-chat prompt into Claude, Cursor, or any MCP-connected agent, or download the SKILL.md file and copy it into your agent's skills directory.

What does the Switch Faces skill do?

switch-faces It is listed under Coding & Dev Tools on SkillMD.

Is Switch Faces safe to use?

This skill has not completed SkillMD's automated safety review yet. SkillMD never runs a skill's scripts for you; review the SKILL.md before installing.

Which AI agents work with Switch Faces?

This skill is tagged as working with Claude Code, Claude.ai, OpenAI Codex. SKILL.md is an open format, so most agents that read a skills directory can load it too.

Is Switch Faces free to use?

Yes. Installing skills from SkillMD is free, and the skill stays under its author's original license.

Who published Switch Faces?

jperrello (@jperrello) published this skill. Their other Agent Skills are listed on their SkillMD profile.

Switch Faces by jperrello — Agent Skill

Name: Switch Faces
Author: jperrello

File contents

switch-faces

When two people share a frame (or a multicam podcast), a real editor cuts away to the listener reacting while the speaker talks — it manufactures "multi-cam" energy and breaks up a static two-shot. This skill does it deterministically: it finds the non-dominant face in the 16:9 source and hard-cuts to a 1080x1920 crop of it during the speaker's natural pauses.

It is the reaction-shot sibling of zoom-punch (re-zooms the speaker at loud words) and jump-cut (reframes the same speaker on a tiling rhythm). All three are deterministic, timeline-preserving lap passes on the clean vertical.

Invoke

.claude/skills/switch-faces/switch-faces.sh <base_vert.mp4> <source16x9.mp4> <transcript.json> <out.mp4>

base_vert.mp4: the 1080x1920 vertical clip to composite onto (post zoom-punch)
source16x9.mp4: the pre-vertical 16:9 source (clip_NN.tight.mp4 / the verify-bookends recut) — holds the listener's face, cropped out of the vertical
transcript.json: the clip-local word-timed transcript (same timeline as the vertical — fill/jump/zoom are all timeline-preserving)

Placement rules

needs ≥2 face identities in the source (dominant speaker + a listener); solo clips → passthrough
listener = a non-dominant face per shot, the calmest (least lip motion) one — reads as listening, not a second speaker
windows = the biggest inter-word pauses (SWITCH_MIN_PAUSE, default 0.32s), biggest first; window starts ~0.4s before the gap so the cut catches the speaker landing the line
hold = ~0.8-1.2s (cycles 1.1/0.9/1.0/0.8), count = clamp(dur/12, 1, 4), ≥5s apart
excludes the cold-open title region (first SWITCH_LEAD=2.8s) and last SWITCH_TAIL=1.5s
a window must lie inside one shot with the listener present for its whole span — never straddles a shot cut
listener framed a touch looser (SWITCH_FACE_FRAC=0.40) than fill-vertical's speaker (0.45)

Chain position

After zoom-punch, BEFORE broll-pick/broll-composite (b-roll cutaways may still override their own windows) and burn-subtitles (captions burn on top of the reaction shot). Timeline-preserving — audio is stream-copied and total duration is identical, so every downstream timestamp stays valid. Orchestrators treat it as non-fatal: on any failure the vertical passes through. Disable with SWITCH_FACES=0.

Output

1080x1920 mp4, video re-encoded, audio stream-copied. Idempotent via <out>.sfmeta. No listener / no pause windows → passthrough copy.

---name: switch-facesdescription: switch-faces---# switch-facesWhen two people share a frame (or a multicam podcast), a real editor cuts away tothe *listener* reacting while the speaker talks — it manufactures "multi-cam"energy and breaks up a static two-shot. This skill does it deterministically: itfinds the non-dominant face in the 16:9 source and hard-cuts to a 1080x1920 cropof it during the speaker's natural pauses.It is the reaction-shot sibling of `zoom-punch` (re-zooms the *speaker* at loudwords) and `jump-cut` (reframes the *same* speaker on a tiling rhythm). All threeare deterministic, timeline-preserving lap passes on the clean vertical.## Invoke```.claude/skills/switch-faces/switch-faces.sh <base_vert.mp4> <source16x9.mp4> <transcript.json> <out.mp4>```- `base_vert.mp4`: the 1080x1920 vertical clip to composite onto (post zoom-punch)- `source16x9.mp4`: the pre-vertical 16:9 source (`clip_NN.tight.mp4` / the verify-bookends recut) — holds the listener's face, cropped out of the vertical- `transcript.json`: the clip-local word-timed transcript (same timeline as the vertical — fill/jump/zoom are all timeline-preserving)## Placement rules- needs **≥2 face identities** in the source (dominant speaker + a listener); solo clips → passthrough- listener = a non-dominant face per shot, the **calmest** (least lip motion) one — reads as listening, not a second speaker- windows = the biggest inter-word **pauses** (`SWITCH_MIN_PAUSE`, default 0.32s), biggest first; window starts ~0.4s before the gap so the cut catches the speaker landing the line- hold = ~0.8-1.2s (cycles 1.1/0.9/1.0/0.8), count = clamp(dur/12, 1, 4), ≥5s apart- excludes the cold-open title region (first `SWITCH_LEAD`=2.8s) and last `SWITCH_TAIL`=1.5s- a window must lie **inside one shot** with the listener present for its whole span — never straddles a shot cut- listener framed a touch looser (`SWITCH_FACE_FRAC`=0.40) than fill-vertical's speaker (0.45)## Chain positionAfter `zoom-punch`, BEFORE `broll-pick`/`broll-composite` (b-roll cutaways maystill override their own windows) and `burn-subtitles` (captions burn on top ofthe reaction shot). Timeline-preserving — audio is stream-copied and totalduration is identical, so every downstream timestamp stays valid. Orchestratorstreat it as non-fatal: on any failure the vertical passes through. Disable with`SWITCH_FACES=0`.## Output1080x1920 mp4, video re-encoded, audio stream-copied. Idempotent via`<out>.sfmeta`. No listener / no pause windows → passthrough copy.

Frequently asked questions

Switch Faces