Skip to content

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

11.8 KB
Raw Blame

World: what a frame is made of

A film is not a layout with animation on it; it is a place with light in it. This file is the richness half of "one focal point inside a built world". Restraint decides where emphasis goes; it does not mean an empty frame. The user said it directly: "It should have more stuff than this, but it should use the animations, the fonts, the logos" (taste/verdicts.md), and v2 builds beat flat silent vector "in everyone" (the 09-30 u3/u6 rounds). The opposite failure is real too: clutter, and everything lit at once. The test is one focal point, many layers.

1. What a frame is made of

Decide these per act, write them in the storyboard (director.md worksheet), and keep them consistent inside the act:

Layer Question Sources, best first
Ground What is behind everything? a surface from the capture (product screen, media plate, signature graphic); a real material (paper, cloth, glass, wood); an imagegen surface plate; a flat colour ONLY if the brand is genuinely flat or it is a flood/dive under ~1 s
Light Where does it come from, what does it do? one key with a direction, a rim to separate the silhouette, falloff into shade; the brand colour as light more often than paint
Depth What is near, what is far? 2-3 planes: a blurred or cropped foreground, the lit subject, a quieter background; parallax in camera moves; depth of field spent once
Texture What does the surface feel like? grain, paper tooth, brushed metal, fabric; a little temporal grain in render (section 5)
Material What are the objects made of? brand plastic (colour + low roughness + clearcoat), satin or anodised, glass, paper, ink; UI with real chrome
Sound What do you hear? room tone or a bed that changes where the picture changes; close dry UI sounds; silence before the hit (sound-sync.md)

A frame that is "a card on a flat colour" has only one of six. A frame with all six and one lit subject is a world. Flat swatch grounds per scene (cream, ink, lavender, green) are a slop pattern: five swatches in 19 s was a v1 test film; it reads as a deck. One light logic and one lens per film: assets in four rendering styles read as stock however good each is.

2. Light logic

  • Key + rim + falloff + contact shadow. One source with a direction, a rim light that separates the silhouette, falloff into darkness (or into a lower-luminance tint on a light ground), and a contact shadow that grounds the object. Flat fills with a drop shadow and a glow on every element feel cheap.
  • A sweep of light travelling across a surface reads as premium; a wide flash over the whole object at once reads as a glitch. Use the sweep on the one thing the idea is about, once.
  • Light only where it means something. The same soft floor glow under every scene reads as a template (taste/edits.md). A glow belongs to an event, a source or an object with a job. A radial glow, vignette and orb around a logo on near-black is the premium pastiche (reach 4).
  • Brand colour as light. A rim light, a horizon glow, the colour of the one light source, a micro-detail amplified into weather (an icon's iridescence becoming the halo of the phone). A gradient works when it is the brand's own material, carries data, or behaves like one real light; it fails as a constant two-colour backdrop.
  • On a light ground the same logic: a soft directional key, real shadows with a falloff, a paper or print texture, a warm or cool cast that matches the key. A light ground is as premium as a dark one; choose by the brand, not by habit (six of ten no-skill AI films sat on near-black, under a third of human films did).
  • Flat brands: if the brand is genuinely flat (a printed, poster-like system), stay flat on purpose: hard colour, hard shadow, grain, and still a ground with an edge and an object with a cause. Say so in NOTE.md.

3. Grounds that are not flat swatches

Build the ground from, in this order:

  1. The capture: film/brand/media/ photography, the signature graphic, a product screen. Crop, push in, blur to a plate, tint, or use as the lit surface. a v1 test film used 0 of 9 captured files and shipped invented UI instead.
  2. A real material the product lives in (paper for a note app, a desk, a counter, fabric), photographed or generated (section 4, imagegen).
  3. A code-built surface that has structure: a shader, a halftone, a paper grain, a lit gradient whose colours come from the product's state colours. Rule: a ground has light, texture or material; the exception is a flood or a dive under ~1 s, or a brand that is flat on purpose. Flip light and dark between acts by a carried object or a hard cut on motion, never by changing swatches for rhythm.

4. Real UI through a camera

"Real UI shown through a camera (cropped, lit, at scale), not a small screen floating in black." Readability comes from the camera, never from inflating the interface: a test film's UI text blown up to 33 px was called a slide; the fix was real size (15-17 pt) and a macro push onto the one element (edits.md).

  • Magnify with the camera: push, snap or crop until the one readable object is the frame (text at least 0.04 H, fragment at least 0.55 W, or cut by the frame edge; this is gate G1).
  • Lit: the UI is an object in the world: it catches the key, has an edge, a surface that contrasts with the ground, a shadow or a rim; it is not a rectangle with a background colour.
  • At scale and framed: crop to the active control, the cursor and its consequence; the whole dashboard small is filler.
  • With identity: the product's real icon, real sender, time, sentence, real status bar or tab bar for a phone, real density (grouped rows, large title), real type. Never the component-library default: a dark rounded card with a status dot and "label . $value", a white card with "+ Aa :)" toolbar circles, a "Good morning" greeting, zinc greys, skeleton bars. Never invented names or numbers where the capture has real ones. lint.mjs flags several of these; the rest you find on the sheet.
  • If the capture has real screens, crop those. If it does not, build the fragment from the product's own tokens, radius and type (see runtime/native-ui/ for true-proportion phone and notification builders) and declare in NOTE.md that it is a rebuilt fragment.

5. Tools and when to reach for them

hero3d (real 3D). WHEN the idea is physical: an object that transforms, catches light, has mass, or that the brand makes (a bottle, a key, a phone, an extruded mark). Not as premium garnish. Use runtime/hero3d (usage in runtime/README.md, section "3D", driven from the film as M.hero3d('#gl', { envMap, objects: [...] | build: (THREE, scene, camera, ctx) => ..., state: (t, objs, ctx) => ({ camera, objects }) }), t in seconds; reference examples/3d-hero/index.html); every frame is a pure function of t.

  • The environment is the lighting: metals and glass show their surroundings, so give them a studio environment (bright cards on a dark field, structure behind the camera). Without an environment map everything is grey CG plastic.
  • One key, a rim, falloff, contact shadows. Materials: brand plastic = brand colour, low roughness, clearcoat; satin = some metalness, mid roughness; glass needs something in the 3D scene behind it to refract; iridescence needs a non-white base.
  • Extruded chrome type with flat faces flashes whole letters white/black; bend the shading normals slightly so each face carries a band of environment.
  • Filmic tone mapping for dark chrome; neutral when a brand colour must be exact. Bloom only on true light sources, low strength, high threshold.
  • Camera like a lens: narrow FOV (about 25-35 degrees), moves with weight, depth of field spent on one rack focus.
  • Slop 3D: default grey material, a spinning logo for no reason, orbiting an object when the film isn't about its form, 3D on every beat, bloom haze, lens flares, glossy clip-art icons (shield, trophy, rocket, lightbulb).

imagegen surface plates (scripts/imagegen.py, Codex image tool; flags in references/tools.md). WHEN a surface the brand doesn't have a photo of is missing: a photographic plate, a texture, a material, an environment, a physical prop. Never text, logos, UI or people presented as real; code owns every letter and every interface. Inspect each image at full size (text leaks, anatomy, style drift) and check the md5 against the previous take (a stale re-use is the usual failure). A plate must PERFORM: be revealed, travelled through, lit, cut out; not sit behind a centred headline. Illustrative imagery is declared in NOTE.md, never captioned on screen. One judged music-app round picked the build with real-looking generated material ("C is very good") because it carried prompt to feed to logo.

shutter blur (node scripts/render.mjs ... --shutter 180 --subframes 8). Code is perfectly sharp and frozen between frames, and the eye reads that as "computer". Shutter blur integrates sub-frames in one pass (4 sub-frames leave double outlines on fast moves; 8 is the floor). Use it for SMOOTH-clock films. One blur grammar per film: shutter blur, OR per-element directional smear, OR crisp steps (twos clock, hard steps). Never stack real blur on fake smear copies, and never use --shutter on a stepped film.

grain (--grain 0.03). A little temporal grain kills banding in gradients and glows and makes flat colour feel photographed. Skip it when the brand is crisp and flat on purpose; skip it on tiny details that must stay clean.

6. The screenshot-worthy hero moment

Before building: name the one frame someone would screenshot and send to a friend. It is a single frame with a focal point, a source of light, a surface you can feel and a thing happening. If you cannot describe it in a sentence, the hero moment is not built; the idea is a topic. The hero moment gets the hardest work: 3D, a shader, a dense composition, a physical simulation, an image plate. It is built FIRST (step 6 in SKILL.md), rendered with --still, and looked at. If it doesn't read at speed, fix the idea or the staging before the rest.

7. Richness is not clutter

  • One focal point inside a built world: many layers, one lit thing. Hierarchy by light and scale, then colour, then isolation.
  • Density that rewards a second look (a plate with grain, a rim, a shadow, a detail in the UI) is richness. Everything the same size and brightness is clutter (and a slide).
  • The user picked the calm legible film over a frantic one: richness is not energy. Check the human band in look.py (mean motion median 6.8, IQR 4.4-9.8) as a question, not a target.
  • Empty is allowed when it does a job: a breath before a payoff, isolation after a matched cut, under ~1.5 s. Otherwise fill the ground with the world.
  • Test: cover the type. Does each frame still show a picture (a surface, light, an object)? If the answer is "a flat colour" or "a card on a flat colour", the beat has no picture.

Slop tells for this topic

AI default Decision
A flat swatch per scene A surface from the capture or a real material; flat only for a flood under ~1 s
Radial glow, vignette, orbs around the logo Light with a source and a direction, at an event
Ghost text at 5-8% opacity, hairline rules, grid patterns, "grain overlay" as decoration (house-style defaults of the HyperFrames skills) Texture as a surface that belongs to the world; this skill wins where they disagree
3D tilt or 3D bar charts around a flat card A cropped, held-pose device and a camera move, or a real 3D object with an environment
A small card floating on a void The fragment magnified by the camera in a lit world
Glass cards stacked over a gradient One material, one light logic
A generated image as a centred backdrop A plate that performs (revealed, travelled, lit)
Bloom and lens flare as "premium" Bloom on true light sources only, or none
Four rendering styles in one frame One light, one lens, one drawing style