Parrass/mograf-public

motion-review: the taste gate

Linters catch broken code. This skill catches taste. Default posture: flag first; approval is earned. A piece that "looks fine" scores 3, and 3 does not ship.

How to run it

• Run as a fresh sub-agent when you can. Give it: this file, the project folder, and the brief. Do not tell it how hard the build was.
• The reviewer never edits the composition. It writes review.md in the project folder and returns the verdict.
• Look at pixels, not just code. Every score must point at a timestamp, frame or selector.
• One reviewer per piece. Two reviewers running at once write the same review.md and their verdicts conflict. Each round ends with one verdict: line; the caller polls review.md for it. Re-reviews go to the same reviewer; start a replacement only if review.md shows no new round after a long wait, note the cancellation in review.md, and have the replacement read the history first. If rounds conflict, the last write to review.md is authoritative.

Procedure

Run every step. Commands assume a HyperFrames project (npx hyperframes …); for other stacks use the equivalent (Remotion: npx remotion still for frames, npx remotion render for the strip).

1. Structure. npx hyperframes check . Any error is a Block. Read the warnings too. Its contrast audit is the contrast gate. Decorative text never meant to be read (blur ghosts, ghost-trail clones, dissolving chips, scramble filler) is excluded with data-layout-ignore on each node; never use it to hide real copy.

2. Motion map. Read the timeline code (or the animation map, if your HyperFrames skills ship animation-map.mjs). Note:

   • dead zones of 1 s or more with nothing moving;
   • spatial moves shorter than 0.2 s (too fast to read) or longer than 2 s (drifting);
   • irregular or mechanical staggers;
   • elements that leave the frame or collide by accident.

   Motion driven through a proxy/setter object (custom camera, three.js, counters) is invisible to tween-based maps; use frame strips (step 5) for those moves.

3. Contact sheet. From the storyboard or beat list, take scene and seam times. Snapshot:

   • frame 0, the hero moment, and the final frame;
   • 25 / 50 / 75 % of every scene;
   • every seam at −2 frames, 0 and +2 frames.

   npx hyperframes snapshot --at <t1,t2,…>. Pass exact frame times (t = frame / fps, e.g. frame 316 at 30 fps → 10.5333); --at does not snap, so a rounded time shows an in-between pose the render never contains. snapshot leaves stale PNGs from earlier runs, so move the old snapshots/ dir aside before each run (e.g. node -e "require('fs').renameSync('snapshots','snapshots-'+Date.now())") rather than deleting with an rm glob; --at can also add a stray frame you didn't ask for, so ignore extras. Open the images and look. Zoom on type only at times the element is visible (zooming on a hidden element can hang).

4. Depth proof. For camera or hero moves, npx hyperframes keyframes --ghost --angle. Planes separating along z mean real depth; everything scaling together means fake depth. Even ghost spacing on a hero move means linear easing, which fails. If --ghost cannot run, take a scratch copy and give the world's parent a proof camera (push back + pitch + yaw, e.g. translateZ(-900px) rotateX(-28deg) rotateY(64deg); yaw alone puts the camera inside deep scenes), snapshot 4–6 frame times across the move and overlay them (ffmpeg … -filter_complex blend=all_mode=lighten). Build this onion by hand for nested perspective (snapshot --angle side flattens it) and for JS projection cameras (3D projected onto one SVG/canvas): swap a side camera into the projection function in the scratch copy.

5. Real-speed watch. If a render exists:

   • make a strip: ffmpeg -i renders/video.mp4 -vf fps=6,scale=360:-1,tile=6x4 strip.jpg;
   • check duration and the first and last frames with ffprobe;
   • if there is audio, check each hit against the cue sheet (±1 frame) and true peak ≤ −1 dBTP on the final AAC file (ffmpeg -i … -af ebur128=peak=true -f null -);
   • transparent .webm overlays: ffprobe shows VP9-alpha as yuv420p; verify alpha by decoding with -c:v libvpx-vp9 before -i (stream reads yuva420p) or the ALPHA_MODE=1 tag.

6. Code read. Grep the composition for:

   • ease: (list every ease used) and durations;
   • opacity-only entrances and small y: offsets (20–40 px);
   • Math.random / Date.now (non-deterministic);
   • back.out|elastic|bounce, yoyo, repeat: -1;
   • linear or power1 on spatial moves; eases outside the allowed list (mg.*, power3.out, expo.out, power3.in, power4.out, power2.inOut, none, a baked spring);
   • concurrent tweens sharing an ease (fine only as a locked group: one multi-target tween or a child timeline with defaults; a setter-driven proxy counts once, by the ease on its tween);
   • a blurred line-mask rise starting below yPercent 130 (blur leaks into the mask);
   • banned fonts;
   • missing grain, vignette or motion blur.

7. Score the 9 dimensions (anchors below), 1–5, each with a one-line reason tied to a time or frame.

8. Craft bar (12 items, pass/fail; table below).

9. Anti-slop scan (top-15 list below). Record ids.

10. Licenses. Every non-original asset (font, SFX, music, image, third-party code) has a source and a compatible license written down. Fonts: OFL or similar. Code: MIT or Apache with headers kept. Anything unknown: flag it.

Rubric anchors (score 1–5)

Dimension	1	3	5
Concept	No idea; text on backgrounds; any brand could be swapped in	Clear one-line idea; motion supports it in some scenes	One strong idea end to end with a surprising, ownable moment you can describe in a sentence
Composition	Everything centered, uniform sizes, clutter or empty without intent	Grid, one hero per frame, decent negative space	Every still is poster-grade: deliberate crops, depth layers, negative space; layouts evolve
Typography	System or generic font, default tracking, whole blocks fading	Right split unit (line/word/char), masked reveals, legible	Type is the animation: variable axes, kinetic layout, crop/scale play, flawless rhythm
Timing & easing	Linear or default eases, uniform durations, no holds	Custom eases on hero moves, durations vary, holds readable	Asymmetric velocity, tasteful overshoot, holds that feel inevitable frame by frame
Choreography	Everything moves together or in a uniform stagger	Hero leads, supports follow with offsets, overlap present	Conducted: one gesture flows into the next, secondary motion reacts to primary
Transitions	Crossfades or unrelated cuts	Designed masks/wipes, on beat	One continuous camera/idea; memorable, never gratuitous
Depth & texture	Flat vectors on a flat or blob gradient, static frame	Camera drift/push, 2+ depth planes, some grain or light	Real 3D or convincing 2.5D, lighting tells the story, texture unifies without showing off
Sound sync	Silent or unrelated audio	Major cuts land on beats	Picture and sound composed together; silence used as an accent; clean mix
Finish	Bugs, jitter, pops, aliasing, unloaded assets	Clean, consistent, designed end frame	Studio-deliverable; every frame scrutinized; end frame held and resolved

Scores 2 and 4 sit between the anchors. To score 4:

• Timing: ease palette by role, "fast out, long settle" arrivals, accelerating exits.
• Choreography: anticipation, follow-through and eased or distance-based staggers.
• Transitions: at least one match cut or morph, and cuts at peak velocity with blur.
• Depth: 3-plane parallax and motion blur on fast moves.
• Composition: at least 5× scale contrast.

A muted, SFX-free piece that declares itself silent in the brief scores Sound sync on its silences and visual rhythm; do not punish a deliberate choice.

Hard gates (any one = Block)

• check fails, the render is non-deterministic, or a seek shows a glitch.
• Text you can't read:
  • hold shorter than 0.5 s + 0.3 s per word;
  • contrast below 4.5:1;
  • text crossing the safe area.
• Font flash, missing asset, broken or empty frame at a seam, or a pop on the first or last frame.
• 3 or more anti-slop hits without a written justification.
• Linear or default ease on a hero spatial move.
• A craft-bar item missing without a written waiver in the brief.
• More than 3 flashes per second.
• A third-party asset with no source or license on record.

Craft bar (12 items)

1. Concept. A one-sentence idea, and a signature moment by ~40 % of the runtime.
2. Hero still. A designed hero frame (Composition ≥ 4, at least 3 depth layers).
3. Camera. At least one real camera or depth move, proven by the ghost or side-angle check.
4. Custom eases. At least 3 distinct ease characters, and no two concurrent tweens sharing one by accident.
5. No default fade-up. No opacity 0→1 + y 20–40→0 anywhere.
6. Named seams. Seams are named and velocity-matched; at most one plain crossfade.
7. Motion blur on 1–3 fast beats (never on text while it is being read).
8. Texture pass. Grain, vignette and radial (light-motivated) gradients.
9. Light event. At least one: a sweep, gloss, scan band, bloom or leak.
10. Sound. Sound sync if there is audio; otherwise the piece is declared silent.
11. Rhythm.
    • The slowest storyboard beat is at least 3× the fastest (beats are storyboard rows, not sub-tweens).
    • Across all tweens, max/min duration is at least 4.
    • The final still holds for at least 0.8 s.
    • The last frame works as a thumbnail.
12. Rubric bar met (see Ship bar).

Anti-slop top 15

id	Tell	Fix
S1	Every element fades in and slides up 20–40 px	Entrance by role: masks/splits for text, scale-from-point for shapes, camera-carried for scenes
S2	Identical durations (all 0.5 s / 1 s)	Duration from distance and mass; hero arrivals 0.8–1.5 s, exits ~30 % faster
S3	Default ease (power1, CSS ease, ease-in-out) or linear on spatial moves	Named custom eases; asymmetric "fast out, long settle"
S4	Uniform 0.1 s stagger	Eased stagger from a meaningful origin (center, focal point, reading order)
S5	Everything moves at once, same speed	One lead; supports follow 2–8 frames later with less amplitude
S6	No holds: in, then straight out	Read hold ≥ 0.5 s + 0.3 s/word
S7	Dead stops: arrives and freezes	Long settle, micro drift through the hold
S8	Elastic/back/bounce on everything	At most one overshoot per piece, ≤ 6 %, on transforms only
S9	Idle "breathing" pulses (scale/opacity yoyo)	Stillness, or one motivated ambient, or a camera drift with an end pose
S10	Fast moves with no motion blur	Shutter-style or directional blur on the 1–3 fastest beats
S11	Centered stack: title / subtitle / button	Asymmetric grid, edge-anchored type, 30–50 % negative space
S12	Static frame, no camera	Push, drift, parallax planes, punch-ins on beats
S13	"Slides": scene fades out, next fades in	Motivated seams: match cut, momentum, mask, morph
S14	Purple-blue / pink-orange blobs, glassmorphism, neon on everything	Gradients that behave like light from a source; 2–3 hues + neutrals, accent ≤ 10 %
S15	Inter/Roboto/Poppins regular as "the design"; emoji or stock icons	Expressive display face, tight tracking, weight contrast; custom marks or none

Also watch for:

• typewriter + cursor as the whole idea;
• generic counting numbers;
• lorem-like copy ("unlock the future of…");
• an ending that just stops or fades out;
• a flat energy curve with no climax;
• motion that would fit any message.

Ship bar

• APPROVE: average ≥ 4.0, no dimension below 3, zero hard gates.
• POLISH: average 3.5–3.9, or one dimension at 3. Fix list, then re-review.
• REWORK: average below 3.5. Go back to concept and styleframes; polishing won't save it.

review.md format

piece: <name> · reviewed: <date> · render: <path or none>
concept: N — …            composition: N — …        typography: N — …
timing_easing: N — …      choreography: N — …       transitions: N — …
depth_texture: N — …      sound_sync: N — …         finish: N — …
average: N.N
hard_gates: [ … ]
craft_bar: [1 ✓, 2 ✗ (reason), …]
anti_slop_hits: [S1, S7, …]
verdict: APPROVE | POLISH | REWORK   → route fixes to: concept | composition | motion build
contact_sheet: <path>

Then a Before | After | Why table. Order it by leverage (impact ÷ effort), and make every "After" an exact value:

#	Where (t / frame / selector)	Before	After	Why
1	3.40–3.90 s #title	opacity + y 30, power2.out 0.5 s	split into lines, mask reveal, settle ease cubic-bezier(0,.65,.51,.99) 0.9 s, lines +4 frames	default fade-up (S1); no settle (S7)

Fix order: delete → reduce → fix easing → fix composition/origin → retime → add depth/blur/texture → polish. Never write "make it more dynamic"; always give a recipe with numbers.

Re-review

Re-reviews go back to the same reviewer. After fixes, re-run step 1, step 3 (changed regions and their seams) and step 7. Any retime re-runs check: its contrast samples move with the timing. Keep the old scores in review.md so the improvement is visible, and end each round with its own verdict: line. Stop after 3 polish loops; if the average is still below 4.0, return REWORK.

---

Part of the free mograf skills (MIT). For the full studio pipeline, see mograf.si. Each piece in the library ships with its prompt, skills, source and copy.