Creative video editor (cinematic)
The opposite of the fast talking-head Reel: dark, deliberate, music-led, every frame composed. The reference is Gawx
Art's "Perfect Imperfection." (studied in references/gawx-cinematic-doc.md, with measured colour, cut rhythm,
typography and structure). Read that file before every project.
What this skill borrows (read-only) and what it owns
Borrowed from video-editor (don't edit it from here) | Owned here |
|---|---|
transcription + EDL + per-segment render: ~/.claude/skills/video-editor/engine/ ($VU_PY $VU/helpers/render.py, run with --draft for 720p review, --uhd for 4K, grade none so the look is applied once, here) | luts/Cine_Gawx.cube, luts/AppleLog_Cine_Gawx.cube (rebuild: scripts/make_gawx_lut.py) |
the sound library ~/.claude/skills/video-editor/shared/sfx/ (manifest.json peak times) | scripts/cine-finish.py: grade + halation + grain + vignette + frame + chapter cards |
fonts in ~/.claude/skills/video-editor/presets/instagram/assets/fonts/ | references/cinematic-storytelling.md: the bookend structure |
| HyperFrames for anything animated | this file's rules |
VE=~/.claude/skills/video-editor; VU="$VE/engine"; VU_PY="$VE/.venv/bin/python"; CV=~/.claude/skills/creative-video-editor
Workflow
- Intake: the story (what was made / where he went / what changed), the platform (Reel 9:16 by default, YouTube
16:9 with
--landscape), the frame shape, music (a calm cinematic or lo-fi track, or he adds trending audio in-app). - Shot plan (before filming, if he hasn't shot yet): dark room + one practical lamp or window on one side of the face; red objects in frame; macro shots of hands and objects; an overhead table shot; a silhouette in a doorway; wide static establishing frames; 2-3 handheld or fisheye selfies; golden hour exteriors. Shoot 4K (Apple Log if he wants the most grading room).
- Structure: the bookend (
references/cinematic-storytelling.md): repeating-pattern opening that reveals what's missing → origin through objects, numbered → a date that sets stakes → a question → process / journey (one music-only montage) → meaning said out loud → the bookend line, completed. Number the chapters. - Cut: calm narrated stretches (a shot every 3-5 s, holds on faces and objects) alternating with music-only montage bursts (a cut every 0.6-1.2 s on the beat). Measured on the reference: average shot 3.0 s, median 1.8 s. Let shots breathe; no punch-ins, no zoom-bounces, no one-word captions.
- Finish:
python3 $CV/scripts/cine-finish.py base.mp4 review.mp4 --res 720 --frame 4x5 --cards cards.json(review), then the same with--res 4kafter approval (render ladder: 720p review, 4K final). - Sound: music bed + real ambience (trains, streets, room tone); sounds only where the picture makes them
(
transition/camera-shutteron photos,cinematic/remembering-wooshon memories,action/*foley on hands working). Master to -14 LUFS, music ducked under narration. - Captions: small and low, pale yellow (
ig-captions.py --style doc-yellowfrom video-editor) or none; chapter cards carry the structure.
Music: ~/.claude/skills/video-editor/shared/music/ (README "Pick by context": cinematic, deep-reflective,
memories-nostalgia, no-talking-storytelling suit this skill; music-cut.py cuts the most-replayed window).
Motion graphics
Same rules and learning loop: ~/.claude/skills/video-editor/shared/references/motion-graphics.md. In this skill motion
stays minimal: slow drifts, chapter cards, words spread across the frame; no bouncy UI.
Cuts (the cinematic grammar)
Read ~/.claude/skills/video-editor/shared/references/cuts.md. This skill leans on: match cuts between chapters and
objects (run ~/.claude/skills/video-editor/shared/tools/match-finder.py on the B-roll; graphic = same shape/framing,
movement = same direction), J-cuts on ambience (hear the train before you see it) and L-cuts of narration over
detail shots (split-edits.py), dissolves for time passing and fades to black between parts
(transition-join.py), cut on action inside process montages, cross-cutting for "preparation vs the day",
and one smash cut from a loud montage to silence before the meaning line.
Look settings (cine-finish.py)
| Flag | Default | Notes |
|---|---|---|
--lut | luts/Cine_Gawx.cube | AppleLog_Cine_Gawx.cube for Apple Log footage |
--strength | 0.85 | 0.6-0.7 for dim selfie footage (keeps the face readable) |
--grain | 6 | 4 subtle, 10 heavy |
--halation | 0.25 | warm glow around highlights (lamps, windows); 0 off |
--vignette | 0.6 | 0 off |
--frame | full | 4x5 (Reel with a picture on black + card below), 137 (the reference's 1.37:1), 239 (scope) |
--cards | — | JSON list of {text, start, end, pos}; small red Inter caps, fade 0.3 s |
Honesty rule: the look is half lighting. Flat, evenly lit footage graded with Cine_Gawx gets the colour but not the mood; say so and suggest relighting (one side of the face, rest dark) or using the darker takes.
Typography (from the reference)
Small red uppercase chapter cards ("1. THE MOVE", "3. THE LAYOUT"), words spread across a wide frame ("SAN · FRANCISCO · 2026"), stacked part titles ("PART TWO / THE BUILD"), credits as a word grid, "THE END" in white + red. Never big headlines; never more than one line of text on screen at a time.
Learning
"Save this to the skill" → one rule here or in the reference file. A new cinematic reference gets its own breakdown in
references/ and, if its colour differs, its own LUT generator in scripts/.