iamajmeena/pro-motion-editor-skill

Pro-Motion Editor (glyph-motions)

onetake (Patrick, feitangyuan/onetake) is the engine; glyph-motions is the taste. Read RULES.md first — it is the curated list of rules that came from real corrections. Then TECHNIQUES.md, GOTCHAS.md. CORRECTIONS.md and REFERENCES.md are the evidence logs.

Layout of this skill

• lib/motion.js, lib/ui_kit.js — onetake's move library (window.OM). lib/glyph.js — ours (window.GM): cameraToClicks, applyTilt, glass, formBox, chip, words, wordList, rollTo, frameIndex, motionBudget.
• scripts/ — onetake's render.py (frame-by-frame + shutter blur), stills.py, probe.py, verify_promo.py, analyze_ref.py, look.py, sfx_palette.py, vo_tools.py; ours: vstills.py (vertical contact sheet; stills.py is 16:9 only).
• templates/reel9x16/comp.html — vertical starter built on lib/glyph.js (test render passed); templates/comp.html — onetake landscape DOM example. references/*.md — onetake's docs (composition, motion-library, rhythm, sound, product-demo, reference-deconstruction).
• examples/ — real builds (code only, no media): juno/ (patching a ChatGPT film to v2 with camera+glass+words), ddr/ (series ident with real-3D icons + reel comp), screenlark/ (shared kit sl.js, 3 concepts, morph-chain ad B2, UI-rebuild Q1).

Protocol (in order, no skipping)

1. Read RULES.md + GOTCHAS.md. New reel? Transcribe immediately (rule 5), show cleaned transcript + timestamps.
2. Beat-by-beat plan (layout per range: face / split / fullscreen / recording; voice-synced). Ask everything unclear in ONE message.
3. Reference deconstruction (analyze_ref.py) -> 3 distinct concepts -> user picks.
4. Beat sheet with CARRY (one container morphs through beats; nothing just replaces).
5. stills (use vstills.py for 9:16) -> 5-9 s trial -> light preview copy -> approval -> next beat.
6. Full render 1080p -> approval -> 4K. Log every correction/learning into this folder immediately.

Non-negotiables (short form; full list in RULES.md)

• Everything is a pure function of time; real video only as jpg sequence; never <video>.
• Camera travels to every click; perspective tilts differently each beat; shutter blur on fast moves.
• Glass boxes form first, items appear one by one, text is written on the spoken word.
• Hook = most advanced part. Variety of layouts. Gradient-only moving background (no particles). Nothing on the face.
• Safe zone x 40-940, y 255-1295. Original UIs only. Do not touch the audio; no SFX unless asked.

New video? (install -> edit)

INSTALL.md installs this skill; WORKFLOW.md is the step-by-step. One-liners: scripts/new_reel.py <Name> <source> (frames + audio + transcript), scripts/transcribe.py (no-VAD, auto re-chunks stretched words/gaps into flags.txt), scripts/make_preview.py (preview / GIF), scripts/vstills.py (vertical contact sheet). Landscape/YouTube: same loop at 1920x1080 (see WORKFLOW.md); the first long-form video will refine it.

Minimal use of lib/glyph.js

const { keys, moves } = GM.cameraToClicks([{ at: 42.62, x: 786, y: 240 }]);
const cam = OM.camera(t, keys, { W: 1080, H: 1920 });
GM.applyTilt(canvas, cam, beatIndex, k);
GM.formBox(g, t, 36.55, 37.4, {x:395,y:226,w:110,h:64,r:32}, {x:54,y:184,w:792,h:115,r:24});
GM.words(g, t, GM.wordList(transcriptWords, 40, 4), 83, 253, 32, 450, '#fff');
window.__motion = GM.motionBudget(moves);