Imported from ohad6k/emulo (skills/video/SKILL.md). Install upstream with npx skills add ohad6k/emulo --skill video. Copyright stays with the author.

Emulo video

You take a brief and hand back a film. Not a prompt, a film: the prompt that built it, the composition files, and frames somebody can look at.

0. Load the profile before you write a word

1. Locate emulo.py two directories above this skill; fall back to ./emulo.py only for a direct repo checkout.
2. Store the resolved absolute path as EMULO_PY, then run python "$EMULO_PY" plugin profile-path --domain video.
3. If it exits nonzero, give its exact recovery instruction and stop loading personal context. Never substitute a generic creative persona.
4. Read every returned path completely. First is the core working profile, second is the video profile. Both outrank anything below.
5. If the repo has videos/README.md, it routes short-form work. This skill owns long-form and commercial film only. A vertical Reel that is a designed film, not a screen recording, is this skill.

1. What you hand back

A folder the person can upload to Claude Design and get a film. That is the deliverable. Not a prompt in a chat message, not a description of a film, not a list of ideas. A folder.

Build it at videos/_kits/<brand>/ in this exact shape, because the shape is what makes it work:

Optionally, and only after the four above are finished: build/piece.jsx plus build/frames/, the film built and rendered locally so you can look at it before anyone else does. That is verification, not the deliverable. A kit with no local build still ships. A local build with no kit does not.

The three things that decide whether the folder works

Ship the engine inside ATTACH/. animations-v3.jsx goes in the folder with the assets. Claude Design needs the runtime it is being asked to write against. Leaving it out is the most common way a good prompt produces a bad film.

Isolated parts, never a screenshot of a whole screen. One object per file, transparent, 1600px minimum on the long edge, plus 4K plates at the film's aspect ratio. A whole screen dropped in is a slab and reads as broken. Claude Design films fail on missing assets, never on prompt length.

Put the instruction in the filename. REF-VERCEL_03_object-became-the-mark.png teaches; ref3.png does not. The person uploading them keeps the names, so every file carries its own lesson into the context.

Name what is missing. Never invent a substitute.

If an asset does not exist, say so in PROMPT.md under its own heading, and say what to do instead. A brief that admits two missing assets is usable. A brief that quietly fakes them produces a film with two broken beats and no way to tell why.

Every prompt contains, in this order and by name

1. the brief in the engine's language: OM_SCENES, CUES, named Easing
2. the transformation chain, as match anchors: what becomes what
3. the asset list, with real filenames, so no beat is blocked on a missing part
4. the failure conditions for this specific film, not generic ones
5. the verification commands, so the claim "it works" has an output behind it

A worked example of the whole shape is videos/_kits/claude/: brief, upload order, 31 assets with the official mark split into stackable layers, twelve instruction-named frames, and two missing assets named rather than faked.

2. The engine

animations-v3.jsx is the Claude Design runtime. Read videos/_kits/_engine/README.md once, then work from this table.

Never edit animations-v3.jsx. Its own header says re-running copy_starter_component overwrites it. Every improvement made there is destroyed on the next host copy. The leverage is in the brief.

The model in one sentence: the animation is one element tree rendered as a pure function of one authored time axis, so nothing mounts or unmounts at a section boundary and any object can persist, move or morph across the whole film by ordinary interpolation. That is why this engine suits the work. Zero cuts, one object transforming across the entire film, and the collision law are its defaults instead of things a brief has to fight for.

<CompositionStage width={1080} height={1920} bg="#161309"
                  scenes={window.OM_SCENES} playback={window.OM_PLAYBACK}>
  <Piece />          // ONE component, the whole film
</CompositionStage>

useComposition() -> { T, CUES, time, duration, authoredTotal, playing }

OM_SCENES and OM_PLAYBACK are JSON string literals in plain inline <script> tags of the main document, not type="text/babel", passed through untouched. Every entry needs a desc, one plain sentence, because the user reads it in the timeline popover. Never hand-set nat.

<script>window.OM_SCENES = '[{"name":"Strike","dur":2.5,"desc":"The mark lands and throws green across the ground"}]';</script>
<script>window.OM_PLAYBACK = '{"mode":"loop"}';</script>

Easing, the complete list. linear; easeIn/easeOut/easeInOut for Quad, Cubic, Quart, Expo, Sine; easeInBack, easeOutBack, easeInOutBack; easeOutElastic. Back and elastic overshoot, so they are the game-feel curves and belong nowhere near a logo or an official brand file.

Render from T only. The exporter seeks each frame with a synchronous commit and may serialise the stage the moment the seek returns, so anything painted from an effect or a private animation loop exports stale.

Never render <Captions>. Ohad films the export off his phone and adds his own Hebrew and English. A burned caption ruins that. Locked on-screen copy is a normal element placed where the composition wants it.

Set bg to the film's ground, never the #0b0b0e default, and never #000000. A designed world needs somewhere for light to fall off to.

Define exactly three or four motion helpers up front and use no easing or transform outside them:

MOTION.slam    Easing.easeOutExpo     arrivals and impacts
MOTION.snap    Easing.easeOutBack     chips landing, overshoot capped at 6%
MOTION.drive   Easing.easeInOutQuart  travel, light, drift
MOTION.settle  Easing.easeOutQuart    the final resolve

3. Assign the references. Never average them.

Every reference gets one job and an explicit do-not-copy list. Without the line "do not average the references" the model blends three styles into mush. This is the single cheapest fix in the whole system.

Write it into the prompt in this shape:

Use X as the PRIMARY authority for: pacing, restraint, match anchors, scale
contrast, camera push and pull, depth, the final hold.
Use Y only for: monochrome identity, typography, exact logo, clarity.

Do not copy X's design. Do not introduce X's colour. Do not average the
references.

Read the source. Five films, one engine.

videos/_handoff/_film-sources/ holds the source of five different finished films, all built on the same engine. Read across them. One film teaches you that film; five films on one engine teach you the workflow, which is the thing this skill is actually for.

Supporting them:

How to read them, and this is the point. Do not copy one film. Open at least three and ask what is the same in all of them: how a scene list is declared, where the transformation chain is expressed, how an entrance is timed against the thing it hits, how the last three seconds are built, what never appears. The constants across five films are the workflow. Anything present in only one is that film's costume, and copying it is how you get a film that looks like a tribute instead of a piece of work.

Two ratios that hold across them:

Motion budget, roughly 70/20/10. Seventy percent of the movement is the object transforming, twenty the camera pushing or pulling, ten the camera travelling. When a film feels busy but dead, this has inverted and the camera is doing the work the object should be doing.

Build in stages, and gate between them. Build the first five to eight seconds. Render. Look at a contact sheet. Only then continue. A twenty second film built before anyone looked is a twenty second film nobody can fix, because the defect is in the concept and the concept is now load-bearing.

Also in the repo, for the current job rather than the general workflow: videos/_handoff/vercel/reel/Storyboard.dc.html (the largest storyboard that exists), and videos/_kits/claude/frames/, twelve stills whose filenames carry the instruction.

4. Build the kit before you write the prompt

Films fail on missing assets, not on prompt length. The Vercel film was built from an 8,000 word prompt with a camera scale graph, a match anchor chain, an easing table and a frame by frame shot list, and every frame of it is grey rectangles standing in for text and a hand drawn triangle in place of the real mark. The prompt was not too weak. It handed the model nothing real to put in the frame, and placeholders are the only thing a model can draw with no assets.

Claude Code builds the kit. Claude Design animates it. That split is the pipeline.

1. Measure the reference at 1:1, one file, never a contact sheet. Write down beats, ground colour, mean luminance, mean saturation, and what physically occupies the frame.
2. Brand intake. Official marks live on the brand's own guidelines subdomain, served gzipped, mixed in with their "Please Don't" examples. Render a light and dark contact sheet and look before naming anything.
3. Rebuild the product UI as live DOM carrying real strings, then shoot it at 3x to 5x into flat plates with _tools/shoot.mjs. Interfaces are rebuilt. World art is rendered and composited. Nothing is faked.
4. Cut parts, not screens. A whole screen is a slab, and a slab cannot be animated, only slid. Ship a parts/ folder of isolated transparent objects at 1600px and up: every button, chip, node, bar, tile, badge, icon. Plus 4K landscape plates for the ground.
5. Split the mark into its own layers. An official SVG splits into layers that stack back pixel perfect. Verify it: stacked against the original, the maximum difference on any pixel should be 1 of 255, which is antialiasing. Then the film assembles the mark instead of cutting to it.
6. Optimise. Every file under about 350 KB, the whole kit under about 4 MB, so it survives being carried into a Claude Design project.
7. Open every file and log its defects in KIT.md. Clipped buttons, missing alpha, wrong aspect, white emblems that vanish on a white card. Every defect you found goes in the prompt under KNOWN KIT DEFECTS with what the film must do about it. If you found none, say you opened every file and found none.

Aspect is a real defect class. Rendering 1200x800 chips into a 224x160 box stretches every flag horizontally by 7%, which is subtle enough to survive review and wrong in every frame.

Published artifacts run under a strict CSP that blocks every external host, so nothing in a kit may point at a CDN. Everything is a project file or a data URI.

5. Write OM_SCENES first

It is the outline, and the section names become the cue vocabulary the rest of the brief speaks in. Six sections is a good film. Ten is a slideshow.

The Vercel film, verbatim, as the shape to aim at:

The Blade 4.6 · Aperture 2.0 · Preview 2.0 · Production 2.25 ·
Everywhere 2.25 · Scale 2.25 · Convergence 2.0 · Lockup 1.75

Each desc is one sentence naming a physical event, not a mood. "The triangle becomes an aperture, the camera pushes through it and pulls back into a full product preview" is a desc. "A sense of speed and possibility" is not.

6. The transformation chain, as match anchors

Name the one object that carries the film and forbid breaking it.

Every scene inherits one visible object, shape, line or motion vector from the previous scene. Never: scene fades out, empty background, next scene fades in. The incoming beat is visible while the outgoing is still 25 to 50% visible.

Worked examples from films that shipped:

• Duolingo. One green circle never leaves the screen: Duo's mark, then the lesson node, then the thing every name collapses into, then Duo again.
• Spotify. The track Midnight Signal is printed on the player bar file and on the phone file, so the same song moves bar to desktop to phone to lineup.
• Emulo. The faceless silhouette is on screen at T=0 and in all six sections, and the payoff is the solid robot arriving in front of it so the two together are exactly the shipped mark.

The colour rule that comes with this. An object that will become an official layer corrects to that layer's colour, never to the brand's published one. Duolingo's duo-01-field.svg is filled #50C800 while Feather Green #58CC02 is the guidelines and product value. Correcting to the published green would have stepped colour on the exact frame that must be invisible. The file wins over the guidelines page.

7. The prompt

Fill this. Delete nothing structural. A brief that is long and contains no filenames is the wrong brief.

You are directing a <N> second <vertical|landscape> spec commercial for <BRAND>,
built on the animations-v3 composition engine. Cinematic. Few objects, large, lit.

FORMAT. WRITE IT LITERALLY.

  <CompositionStage width={<W>} height={<H>} bg="<ground hex>"
                    scenes={window.OM_SCENES} playback={window.OM_PLAYBACK}>

<W> wide by <H> tall. If your composition is <wider than it is tall | not
<W>x<H>> you have failed before anything else is judged.

THE ONE LAW

Every object on screen is a file from the list below. The timeline never draws a
<BRAND> screen, never draws the mark, never stands text in with a grey
rectangle, never approximates imagery with a gradient. It places, masks, crops,
lights and moves those files.

If a beat needs an object the kit does not contain, say so and stop. Do not
substitute a div. If you cannot resolve the paths, tell me before building
anything rather than drawing replacements.

REFERENCES

<one job per reference, with a do-not-copy list. Then: do not average them.>

THE ASSETS

  <every file, its pixel size, and one line on what it is for. Name which
   element level parts exist and what they are meant to do on their own.>

THE BAR, AND WHAT IT RULES OUT

<the genre exclusions, not craft notes. Genre exclusions are what the model
already understands: six disconnected slides, a SaaS product tour, a dashboard
montage, a conference presentation, a developer tutorial, a browser screen
recording, a wall of technology logos, a generic futuristic AI advertisement.>

- no paragraphs, no terminal dumps, no reports, no lists of findings. If a frame
  needs more than about eight words to be understood, the frame is wrong
- no charts, no graphs, no dashboards, no code editors
- no neon, no glow, no purple, no gradient meshes, no particle fields

At most ONE object carries text in any frame, and that object is a real thing
with edges, not a caption floating in space.

GROUND AND LIGHT

Measured off the reference film, not chosen:

  frame average sits between <hex> and <hex>, <channel> dominant
  mean luminance <n> of 255
  mean saturation <n>

The base is <hex> and never pure black. One warm key light, <position>, fixed.
Every object carries a contact shadow and a cast shadow. Depth is blur and
scale. Bloom on <named elements> only. One grain pass at low opacity.

LOCKED COPY

  <three or four lines, nothing else>

No feature list, no metric, no URL, no invented tagline, no subtitle under the
logo. Set in <typeface, and say plainly if it stands in for a licensed brand
font>. Revealed through a horizontal mask, never typed letter by letter.

THE ENGINE

<script>window.OM_SCENES = '[
 {"name":"<Name>","dur":<s>,"desc":"<one plain sentence>"},
 ...
]';</script>
<script>window.OM_PLAYBACK = '{"mode":"loop"}';</script>

Key everything to T from useComposition() against those cues. One continuous
composition, zero cuts.

  MOTION.slam    Easing.easeOutExpo     arrivals and impacts
  MOTION.snap    Easing.easeOutBack     landings, overshoot capped at 6%
  MOTION.drive   Easing.easeInOutQuart  travel, light, drift
  MOTION.settle  Easing.easeOutQuart    the final resolve

Nothing eases outside those four. No elastic, no Back curve on an official brand
file, no setTimeout, no requestAnimationFrame, no useEffect painting, no wall
clock, no random, no CSS keyframes running on their own. Render everything from
T so any seek reproduces its frame. Never render <Captions>. Never edit
animations-v3.jsx.

THE ANCHOR

<the one object physically present across beats that carries the film. Name it
and forbid breaking it.>

THE SECTIONS

<one block per OM_SCENES entry. Each block carries:
   what is on screen and which file
   coverage as a percentage of frame width for every hero object
   what causes the next thing to happen
   a "Do not:" line specific to this beat>

CAMERA

<N> events and nothing else moving the frame:

  1. <cue range>   <what>
  2. <cue range>   <what>

Between them the camera holds and the objects do the work. No drifting, no
breathing zoom, no shake, no orbit, no continuous parallax.

Motion budget:
  60% local object and geometric transformation
  20% surface, lighting, parallax, environment
  20% camera scale and reframing

WHERE THE COLOUR COMES FROM

<name the objects allowed to be saturated, and say the ground picks up their
colour. Forbid coloured light with no object making it.>

MOTION SHAPE, MEASURED

<the numbers from §8. State the reference figures and the floor this film must
clear.>

KNOWN KIT DEFECTS

  <every defect you found opening the files, and what the film must do about it>

FAILURE CONDITIONS

<15 to 25 lines, each a specific observable defect for THIS film. §10.>

BEFORE YOU BUILD, RETURN AND STOP

1. the CompositionStage line, so I can see the dimensions
2. the OM_SCENES literal
3. a <N> frame storyboard, and per frame: which asset, its coverage as a
   percentage of frame width, how many elements are moving, and the WORD COUNT
   on screen. Any frame over eight words is a defect, tell me and fix it
4. your impact list with times
5. confirmation the anchor object is on screen in every section
6. anything you cannot verify

Do not animate until I approve.

Then PROMPT-2.md, the approval:

Approved. Before you build, confirm these four:

  <four checks specific to this film, each one a thing that was measured off the
   reference and could quietly go missing>

Now build it from the approved storyboard. Deterministic, keyed to T, kit files
only. When it is running, show me <the cue boundaries> at full size before
anything else.

8. The motion floor, measured

Zero cuts alone reads dead. Cut rate and energy are two separate specifications and a brief has to carry both.

Frame to frame motion energy, sampled at 15fps:

Read the v2 row. Half of that film is frozen. The mean looks acceptable only because a handful of enormous transitions drag it up, and the peak is six times the reference. That is the signature of a slideshow with big transitions between static cards. A motion designer produces the opposite shape: a lower peak and far less stillness, because something is always moving.

The floor to write into a prompt:

• near-still frames below 26%
• no single frame above 25 energy
• frames above 4.0 above 22%
• at least two elements mid-transformation at every moment that is not a hold
• a growing hero gains at least 35% of frame width per 0.5s
• holds are named and counted. Two or three in a film, plus the final lockup

The freeze test. If a frame's motion would survive freezing every gradient in it, the motion is real. A continuously drifting pool or breathing gradient lowers the near-still number without adding one frame of real motion. That is the metric moving, not the film, and we have shipped a bad video that exact way before by optimising the measurement instead of the thing it measured.

Luminance is one film or two. Measured on a rejected build: greetings 19.9, transition 75.1, UI middle 226.2, outro 180.4. A 19.9 to 226.2 span is not a grade change, it is a different film glued between two designed worlds. Name one ground for the whole piece and a ceiling for the brightest section.

Coverage stops things coming out tiny. Hero object 25% to 45% of frame width as a floor, and higher on payoff beats. In the rejected middle the nodes sat at roughly 6%, which is why nothing read and nothing felt like it was moving.

The source material never leaves the middle. Measured on two boards that came back: the product was off screen for 72% and 68% of running time. Both opened on the real product, spent the middle on coloured rectangles, and returned for a closing beat. That is a template reel with a product bookend, and it is the single reason both read as cheap. State a floor as a percentage of frame area and put "any beat takes place on a bare ground" in the failure conditions.

The rule is about chrome, not content. Name the product's own components off the screenshot, the sidebar, the tab strip, the chip row, the player bar, and require them present by name. The strongest form: let the product's own controls cause the film. A chip being struck is what recolours the page. Then collision, causality and real product behaviour are the same move, and the film cannot be described without naming the product.

9. What physically happens in the beat

Everything in section 8 can be green on a film that comes back rejected. The verbatim verdict on one that was: "its just boring words with ok animations instead of showing crazy animation crazy things happening with animation like you did with the hook with the puppet inside the chatbox ui." That build cleared the structure, the coverage floors and the energy floors. It was still a sequence of phrases arriving over a lit ground.

Read videos/_handoff/_film-sources/ITERATION-LOG.md before you write a brief. It is the only file that records what was rejected, in his words, next to what replaced it. Finished code cannot tell you the words were there first.

The per-beat test

For every OM_SCENES entry, write three lines before anything is built. If line 2 is empty, the beat is a card, and a card reads as boring however it eases in.

1. the phrase on screen
2. a body doing something, named with a verb: dives, hammers, hops, is pulled off its feet
3. the surface it does it to, named as a kit file or a rebuilt UI part

A body is a character or an object with mass. A word is not a body. Letters that hammer, shatter or fly are still kinetic typography, which is the exact thing that got rejected. If line 2 names the phrase itself, the test fails.

Rejected: "Create completely" over a fixed box at left: 1090, top: 224, width: 470, height: 630, whose children only widen in place (claude-film.jsx:530-537). Approved: nine blocks fly in from left: tx + Math.cos(a) * 1000 * (1 - p), each carrying blur(6 * (1 - p)) and rotate((1-p) * 26deg) scale(.7 + .3*p) on a 52ms stagger, while a body hammers alongside and squashes on every landing (claude-master.jsx:1012-1044).

Note what separates them, because it is not the count of pieces. Nine flying rectangles with no body beside them is still the rejected beat with more parts.

Ban grow-in-place

Fixed left and top with width, height or opacity animating inside the object's own bounding box is what a model writes by default, because it is safe and can never collide. Every hero entrance carries position, rotation, blur and scale at once, and starts at least 0.35 of stage width outside where it lands. Express every distance as a fraction of stage width, never in raw pixels: the films are authored at 1920x1080 and the rebuilt UI at 2560x1600, so a pixel figure copied between them is wrong by a third.

An object that travelled has a before-state the viewer never saw, so its arrival carries information. An object that grew was always there.

The character must be attached to real geometry

Derive the body's transform from the target object's live animated values every frame, including rotation:

const rad = bars[0].rot * Math.PI / 180;
cx0 = bars[0].x - Math.cos(rad) * bars[0].w / 2;
cy0 = bars[0].y - Math.sin(rad) * bars[0].w / 2 - 34;
rotc = bars[0].rot - 90;

That expression is the tip of the rotated bar, and rot - 90 keeps the body standing perpendicular on it (claude-master.jsx:833-838). Generate paths from the target's own data too, so a climb right to left is an emergent property of the chart rather than an authored path that happens to match.

The reference film also gets this wrong, and you should see where. In the research dive the source cards compute cy = -180 + 1460 * easeInOutCubic(rDive) while the critter three lines below is y = -200 + 1500 * easeInOutCubic(rDive). Two independent expressions off one progress variable, 20px apart at the start and 60px at the end. Delete the critter and every card lights exactly the same. It reads acceptably and it is still a sticker, and it will drift the first time anything is retimed. Couple to the object, not to a shared clock.

Simulated UI, operated at speed

He asked for this by name: "i want like in the reference video when it shows like a mimic ui fast paced doing on it with the mouse and text and stuff." Section 7 bans faking a screen as static art. It does not ban rebuilding one. Never draw a grey rectangle standing in for a product screen. Rebuild the UI as live DOM, one element per part, and operate it. When a film has a product in it, it must carry at least one operated-UI beat.

• Rebuild, do not screenshot. ui.jsx builds sidebar, top bar, greeting, six shortcut tiles and a six-card carousel as separate positioned divs. A screenshot can only be pushed and cropped. A DOM rebuild lets one card lift.
• The cursor is an object with momentum, and every key carries its intent in a comment. film.jsx BALL_RAW is 39 keyframes: { t: 2.80, x: 812, y: 872 } grabs the progress knob, { t: 3.86, x: 960, y: 836 } punches play. A failure is recorded in the same file: easeInOut on consecutive keys made the roll stop and restart at every card. Use one ease-out across the span and let friction decelerate.
• Sample the path, then derive everything from one number. const bPrev = ballAt(T - 0.022); const speed = Math.hypot(vx, vy) feeds squash clamp(speed/5200, 0, .46), heading Math.atan2, motion blur clamp(speed/900, 0, 9), trail opacity and audio gain. Hand-keying five of those separately is how a fast move ends up mismatched.
• Every UI state change is caused by an impact, and the two tables share timestamps: IMP = [0.30, 2.62, 3.86, 4.24, ...] against HOME_TRIG = [4.24, 4.52, 4.80, 5.08, ...], each region a 40ms staggered cascade off its own hit. A screen that assembles on a schedule is a slideshow. A screen that answers 300ms after something hits it is alive.
• One impact function drives shake, camera bump and blur together, sign flipping on alternate hits: Math.pow(1 - d/.34, 2.1) * Math.sin(d*108) * (i % 2 ? 1 : -1), consumed as cam.s *= 1 + sh*.022, shakeX = sh*11, mBlur = min(abs(sh)*4, 7). Fifteen hits in sixteen seconds, none of them feeling repeated.
• A control being operated reads its value off the cursor, it is not keyed alongside it: scrubF = clamp((b.x - TR0)/(TR1 - TR0), 0, 1). Two objects keyed in parallel drift a few pixels and the shot reads fake.
• A click is three asymmetric ramps: 80ms down, 20ms held, 120ms up, with the button swelling 12% starting 160ms before contact. Symmetric press animation reads as a CSS demo.
• Passing near a thing knocks it, by distance not by keyframe: knock = rolling * max(0, 1 - abs(b.x - CARD_CX(i))/120). Hand-authored reaction windows break the moment you retime the roll. A falloff survives it.
• Dim the interface under the copy, never cut away from it. DIM keys brightness across the whole UI, 1.0 down to 0.42 then 0.26, plus a corner scrim. The app stays visibly running behind the words.

Delete these from the background

The fix for a boring beat is more happening in the middle of the frame and less at the edges. Deleted from the approved build: giant background words, the word-storm, decorative hairlines, ambient panels behind the hero, and the light field cut to a single soft wash. STORM is still declared empty at claude-master.jsx:192-193 with its renderer live, which is what deleting it properly looks like.

Occupied holds

Section 12 law 5 fixes the camera to three moves. This is its other half: fill every camera lock with a body doing something. All three locks in the approved film are occupied, and the longest one is the busiest stretch in the picture. When the frame holds, the content earns the motion.

Gate the weakest beat, not only the opening

The eight-second gate in the next section catches a bad opening. It cannot catch a strong opening followed by three flat beats, which is exactly what was rejected here: "the intro very strong... the only parts i dont like is the think create handoff."

So after the opening passes, render the capability stretch and look at that too. Do not self-grade it. Put the contact sheet in front of the person who asked for the film and ask which seconds are the worst. Their answer is the list. Rebuild those beats against the per-beat test above, and only then build the rest.

10. Build it, look at it, then continue

Do not build the whole film and then find out.

1. Build the first 5 to 8 seconds only. Wire build/index.html and build/piece.jsx against the kit. The host scaffold is fixed, and videos/_kits/duolingo/build/ is the working copy to clone:

<script>window.OM_SCENES = '[...]';</script>
<script>window.OM_PLAYBACK = '{"mode":"loop"}';</script>

<script src="../../node_modules/react/umd/react.development.js"></script>
<script src="../../node_modules/react-dom/umd/react-dom.development.js"></script>
<script src="../../node_modules/@babel/standalone/babel.js"></script>
<script type="text/babel" src="../../_engine/animations-v3.jsx"></script>
<script type="text/babel" src="./piece.jsx"></script>

<script type="text/babel">
  /* ?still=1 freezes the transport so the capture harness can seek
     deterministically. The shipped composition autoplays as normal. */
  const still = new URLSearchParams(location.search).has('still');
  ReactDOM.createRoot(document.getElementById('root')).render(
    <CompositionStage width={1080} height={1920} bg="#06120A"
                      autoplay={still ? 'false' : 'true'}
                      loop={still ? 'false' : 'true'}
                      scenes={window.OM_SCENES} playback={window.OM_PLAYBACK}>
      <Piece />
    </CompositionStage>
  );
</script>

piece.jsx opens with one path constant and the four motion helpers, and nothing else in the file knows where the kit lives:

const BASE = '../ATTACH/';   // the ONLY line that changes when assets move

const MOTION = {
  slam:   (o) => animate(Object.assign({ ease: Easing.easeOutExpo },    o)),
  snap:   (o) => animate(Object.assign({ ease: Easing.easeOutBack },    o)),
  drive:  (o) => animate(Object.assign({ ease: Easing.easeInOutQuart }, o)),
  settle: (o) => animate(Object.assign({ ease: Easing.easeOutQuart },   o)),
};

2. Shoot it and look. Snapshots cost seconds, renders cost 25 minutes.

node videos/_kits/_tools/frames.mjs <project> 0,0.6,1.4,2.2,3.0,4.1,5.2,6.4,7.6

The tool prints the stage size and duration, flags a wrong aspect, and prints every page error under PAGE ERRORS:. A single Failed to load resource: net::ERR_FILE_NOT_FOUND means the film is drawing nothing where an asset should be, and it is blocking. Fix BASE or the kit path before you judge a single frame.

The seek listener is on the stage element, not on document; dispatching on document silently does nothing and every frame comes back identical. The stage autoplays, so seek before you shoot. 3. Gate on quality before continuing. Open the frames at 1:1, not as a contact sheet. If the first 8 seconds are not the level, nothing after them will be, and every hour spent on the back half is spent twice. 4. Only then build the rest. 5. Diff the template against the generated file whenever another tool has touched the project.

11. Failure conditions

Generic failure conditions catch nothing. Write 15 to 25 lines, each a specific observable defect for this film. These are the ones that recur, as a starting set to specialise:

• the composition is not <W>x<H>
• any object on screen was drawn rather than placed
• the ground is pure black anywhere
• a frame is empty between beats, or a beat takes place on a bare ground
• any frame contains a paragraph, a report, a terminal dump, or a list
• more than one object carries text in the same frame
• a frame needs more than eight words to be understood
• the anchor object is absent from any section
• the mark or lockup is stretched, skewed, rotated, glowed, recoloured, extruded or animated
• a gradient stands in for real imagery
• the camera moves outside its named events
• the final hold is shorter than 1.5 seconds, or the mark is still moving on the final frame
• the film cuts, or <Captions> is rendered
• animations-v3.jsx was edited
• it reads as a product tour rather than a film

12. Verify, with output

A green command is not visual approval. Report the numbers, do not estimate them.

node videos/_kits/_tools/frames.mjs <project>

stage dimensions, duration, and the contact sheet. Wrong aspect fails here.

ffprobe -v error -select_streams v:0 -show_entries stream=width,height,r_frame_rate,bit_rate,duration -of default=nw=1 <export>.mp4

the floor for a landscape film is 1920x1080, 60fps, 4.69 Mbps.

ffmpeg -i <export>.mp4 -vf "select='gt(scene,0.2)',showinfo" -f null - 2>&1 | grep -c showinfo

cut count. A promo should return zero. scene_score measures nothing else on a zero-cut film, so it is a cut detector and never a motion metric.

Motion energy, near-still percentage, peak and per-section mean luminance are measured off the export, not estimated, and reported next to the floors in §8.

Browser frames carry zero motion blur, so an export reads as stop-motion next to the preview. Render at 240fps and tmix=frames=4 down to 60. Smoothness and sharpness are separate problems: duplicate frames, BT.601 versus BT.709, chroma and pixel ratio are the measured causes, and -tune animation, -tune grain and gradfun are all wrong advice.

For a vertical film the Instagram feed crops the top 285 px, so nothing load bearing goes there.

13. The laws that cost real renders

Each of these was paid for once. None of them is a preference.

1. Reference roles are assigned, never averaged. Vercel for transformation and depth. Spotify for camera and easing. Write "do not average the references" verbatim.
2. Build the first 5 to 8 seconds and gate on quality before continuing.
3. Never fix a concept problem with easing. If a transition reads as a cut, the thing you animate and the thing you land on are two different objects. No curve repairs that. Split the mark into layers and assemble it.
4. A flash cannot replace geometry in a transition. Brightness that hides a missing seam is a defect. Every luminance jump names the object state, evidence or medium change it accompanies.
5. The camera moves once, at the moment that matters. Three events in a 12 second film, and nothing else moving the frame. Between them the camera holds and the objects do the work.
6. Objects must exceed the frame. Everything safely centred reads as scaling, not as camera movement. The player bar at its peak is wider than the frame. The screen fills past its edges.
7. Zero cuts alone reads dead. Energy is specified separately from cut rate, with the numbers in §8.
8. Rendered assets, not browser-drawn shapes. A div is not a card, a gradient is not album art, a triangle you drew is not the mark. If the kit does not have it, stop and say so.
9. Nothing fades in. Every entrance hits something and the impact causes the next beat. A particle that is not consumed by the letter it belongs to is litter, not a cause. Either it causes the thing or it is not there.
10. 2D, never 3D. The cinematic look is flat layer compositing. The camera push is scale() on a container and depth is blur plus scale. 3D scenes are what produce the game look and the lag. Flat compositing is not flat lighting: name a light source with a position, give grounds a falloff, give every object a contact shadow and a cast shadow.
11. Every number on screen is a real published figure, read on a stated date and hardcoded. Never invent one, round one, animate a count up from zero, or turn it into a chart.
12. Do not open a prompt with an essay about why the last attempt failed. These prompts do not apologise or narrate history. They specify. Do not describe a mood where a number will do. Do not write a scene without a coverage percentage.

14. Working files