Skip to content

Looksmaxxing for motion graphics. An agent skill that turns AI slop videos into films that look made by a motion designer.

8.9 KB
Raw Blame

Tools: every script in scripts/

SKILL = the skill folder. All scripts print --help. Zero dependencies unless noted (Node 22+, Chrome, ffmpeg, Python 3.9+). Nothing here calls a paid API except voice.py (ElevenLabs) and, only when asked, imagegen.py (Codex login, no key). Scratch to delete after tests: frame sequences, drafts, look/ dirs.

step script one line
0 providers.sh what this machine has (Chrome, ffmpeg, ElevenLabs key, Codex + image_generation, disk)
1 brand.mjs, fetch_logo.mjs capture a site's brand; fetch an official third-party mark
3 precedent.py rule-only precedent per beat
4 imagegen.py generated SURFACE plates (never text/logo/UI/people)
5, 8 voice.py, sync.mjs, mix.py voice, word timings to frames (5); SFX, music, final mix (8)
6-7, 9 render.mjs, look.py, lint.mjs render, inspect (G0 G2 G3 + audio), page-chrome lint (G5); final render with sound (9)
9 blind_review.py blind A/B page for the user

Gates and who computes them

Same definitions as SKILL.md and slop.md Part 2.

gate computed by note
G0 render exists (decodes, non-blank, moves, right length, audio) look.py --expect S --expect-audio FAIL lines name the cause
G1 proof readable by eye on look/sheet.jpg and stills (no script number) text >= 0.04 H, fragment >= 0.55 W or edge-cropped
G2 no empty frame look.py a run of > 3 flat frames; <= 3 f dips inside a transition ignored
G3 end card short look.py end hold (picture unchanged at the end) <= 1.4 s, i.e. a resolved hold of 0.6-1.4 s; under 0.6 s is a NOTE (ends mid-motion); final shot (last cut to end) <= 25 % of the film
G4 one hero per frame by eye on the sheet (no script number) two panels only if one is >= 2x the other's area or the camera moves between them
G5 no page chrome lint.mjs (GATE G5: PASS/FAIL, exit 1 on FAIL) C1-C4, C6, C7, P1, V1, V2, V5 FAIL; D1 FAILs with media/sub-line/kicker beside it (WARN alone); T1 FAILs when letters grow (a caret alone is WARN); C5, V3, V4 are WARN
look.py writes G0, G2, G3 to look/gates.json and prints them in its GATES block as G0 render exists, G2 no empty frame, G3 end card hold, CTA share (G1, G4 and G5 are named there as not computed by it); lint.mjs prints `GATE G5: PASS FAIL. Quote the numbers in NOTE.md`. A PASS is not good taste; a lint is a lead, not a verdict.

providers.sh

bash "$SKILL/scripts/providers.sh" -> JSON: ready, chrome, ffmpeg (+missing filters), elevenlabs.key, codex.image_generation (true / false / null), disk_free_gb, disk_low. Never installs. Low disk (< 3 GB): short low-res drafts, delete outputs.

brand.mjs

node "$SKILL/scripts/brand.mjs" https://site.com film/brand -> brand.json, BRAND.md, board.png (LOOK at it), logo/, fonts/, screens/, media/. Flags colour roles, blank images, nav-strip screenshots. Empty page = blocked: say so, fall back to fetch_logo.mjs + what the user supplies.

fetch_logo.mjs

node "$SKILL/scripts/fetch_logo.mjs" "<brand name>" <domain> --out film/assets/logo [--slug simple-icons-slug] [--json] -> logo.svg (mark), wordmark.svg, logo-light.svg, logo.json (source URLs, official hex, guidelines link). Order SVGL -> Simple Icons -> the site's own icon/header SVG -> raster icon. Never draws or recolours; if nothing is found, set the name in type. Nominative use only; render the SVG once and look. Exit 1 = no SVG mark.

precedent.py

python3 "$SKILL/scripts/precedent.py" query --beat proof --content ui --text "selection becomes result" --k 6 (beat hook|turn|proof|cta; content ui|type|logo|number|3d; --src rule,hf,lib). Also show PR-A02, rules --beat cta, stats. Returns rules with ids: PR-Xnn measured WHEN/DO/BECAUSE rules (evidence M-ids), PM### the 53 study moments (role, mechanic, why), PL#### rules from 39 other films (opaque film id). In the beat table write the ONE rule you adopt and its id. Rows are model-decoded and unverified: open the cited M-moment before trusting a number; never outrank references/ or taste/verdicts.md. Data: library/moments.jsonl (no copy, frames or brand names).

imagegen.py

python3 "$SKILL/scripts/imagegen.py" plates.json --out film/assets/gen --dry-run first (guard + exact prompts, spends nothing), then without --dry-run (--jobs 3, --only id, --force). One private working dir per plate, md5 collision check, alpha >= 250 snapped to 255, <id>@WxH.png cover crop, ledger.jsonl. The guard refuses (exit 2) any plate that asks for text, logos or brands, UI/screens/charts, people/faces. Code first, captured asset second, generate last; every plate must perform (parallax, light, occlusion, mask, rack focus); label as illustrative; never replaces product proof. Needs codex.image_generation: true; about 1-2 min per plate.

voice.py, sync.mjs, mix.py

  • voice.py subcommands: voices, vo --script script.txt --out audio/vo.mp3 (writes vo.words.json, vo.lines.json; --no-refine skips the audio refinement), vo-local (macOS say, estimated timings), refine, split --vo ... --lines ... --out audio/vo (per-line clips + lines.index.json), sfx --prompt ... --out ... or --batch sfx.json, sfx-norm, music --prompt ... --ms N --out ... [--dry-run], music-check, music-fit --duration N [--loop-bars]. Word and line times are refined against the audio. Reads the ElevenLabs key (exit 3 without one); --dry-run (music) shows the request.
  • node sync.mjs audio/vo.words.json --fps 30 -> beats.json / beats.js: word -> frame table, on-screen onset frames (2 f before the word), clip table for mix.py.
  • python3 mix.py --duration 16 --vo clip@t ... --music m.wav --sfx f@t:gain --duck --out mix.wav -> -14 LUFS / -1 dBTP stereo 48 kHz. Duck is predictable (--duck-db 9).

render.mjs

node "$SKILL/scripts/render.mjs" film/index.html film/final.mp4 --audio audio/mix.wav (any __seek / __timelines / gsap page; --fps, --from/--to, --scale 0.5 for drafts, --still 0.5,2,4 for PNG stills). Writes final.mp4.events.json from window.__events (look.py uses it for exact cuts).

  • Finish, single pass: --shutter 180 --subframes 8 [--grain 0.03 --grain-seed 7]. N sub-frame seeks per frame across the open shutter (forward from the frame time, so a hard cut never ghosts), averaged by ffmpeg tmix inside the same encode; no intermediate files. Cost = N x render time, so use --from/--to for hero ranges. Stepped content (typing, counters, clock:'twos', 1-frame kf steps) stays crisp; smooth motion blurs. ONE blur grammar per film: shutter OR per-element smear OR crisp steps, never shutter over smear. Grain 0.03-0.06 only on gradient/photo/3D films; same seed gives the same bytes. Stills ignore both.
  • WebGL pages: do not mix GPU and software GL in one render; --scale does not cut WebGL cost.

look.py

python3 "$SKILL/scripts/look.py" film/final.mp4 --out film/look --expect 16 --expect-audio -> sheet.jpg (describe each frame in one sentence), cuts.txt, still.txt, mid.jpg (2 f before each cut), strips.jpg, gates.json, and a printed report: cuts (events win), shots, end hold, then the AUDIO block (LUFS, LRA, true peak > -1 dBTP flagged, audible span, silent tail > 0.5 s, dropouts, sound onsets, share of cuts within +-80 ms of an onset vs chance), the GATES block (G0 G2 G3 PASS/FAIL with times) and the MOTION note (mean luma change vs the human band, median 6.8, IQR 4.4-9.8: a question, never a gate). --strict exits 3 on any FAIL. Shutter-blurred mid frames are expected; judge the thumbnail, the cover and the mid-change frame.

lint.mjs

node "$SKILL/scripts/lint.mjs" film/index.html [--times 0.5,2,4] [--samples N] [--json] -> PASS / WARN / FAIL per rule (C1 corner/edge label, C2 tracked caps, C3 counter, C4 metadata, C5 mono label, C6 bar/row, C7 kicker, D1 deck layout, P1 progress bar, T1 letter-typed headline, V1-V5 vibe-coded UI) with sample times, element snippet and bbox, then GATE G5. Each position rule must hold across 0.1 s (a label flying through a corner is not chrome). Text inside a UI surface (card-sized fill + radius, or [data-ui]) is exempt except greetings, "label . value" and the V5 cluster; [data-ui="captured"] (a faithful rebuild of a captured screen) also skips V5 and is printed. Fix by deleting the element; do not justify it. Cannot see text inside <canvas>/images. Exit 1 on FAIL.

blind_review.py

python3 "$SKILL/scripts/blind_review.py" film/review --round "before vs after" v1/final.mp4 v2/final.mp4 (the first argument is the output dir, default review) -> film/review/index.html (random L/R, synced play, frame step, sound per film, pick / slop / 1-10 / notes, Copy verdict), videos hard-linked, key in hidden .answer-key.json (do not open). Offer it when two versions exist; append the pasted verdict to taste/verdicts.md. Judge picture silent first, then each soundtrack alone.