Motion design launch videos

This is how Chan wants launch films made. It is the distilled direction from every prompt and review that produced "The Spark" (v1) and the real-data cut (v2) for Cheerful. The prompts themselves, verbatim and in order, are in references/prompts.md — read them before starting; they are the spec. The code patterns are in references/technique.md. A complete working film to copy from lives in ~/code/cheerful-motion/launch2/.

The stance: you are the director

• No constraints on creativity. Chan says "I want you to be the director", "full creative control", "think out of the box", "no limits on shapes or buttons". Do not ask permission for creative choices; make them, log them as DECISIONs, and list them at the end so he can overturn any in one reply.
• Original, never a template, never a copy. Do not lift the product website's copy, UI, stats or images wholesale. Invent the scenes. "Anything that looks like a template" is banned.
• Launch-film energy. Modern, futuristic, premium — the way top motion designers do it (Linear, Apple, Arc). Something happens on every beat; there is no dead time.
• Attention to detail is the product. Micro-interactions (button press, knob stretch, name pill lag, click ripple, odometer settle, sticker tilt) are what make it read as expensive.

Non-negotiable motion rules

1. One shape, never cut. Every state is the same element morphing its size, radius and colour while its content swaps with a short blur. Cuts are not encouraged; transitions are morphs, camera moves, zooms into an element that becomes the next scene (a tile opens into an inbox, a reply card becomes the order, the cursor becomes the logo star).
2. Content swaps with a short blur, and text inside a morphing container gets its own enter and exit timing — exit short and early — or old and new text overlap.
3. Springs everywhere, a tiny overshoot at most. No bouncy easing. Closed-form damped spring step responses only (see technique).
4. Leading/trailing edges on different springs so things stretch: tab indicators, toggle knobs, liquid pills — the leading edge runs ahead of the trailing one.
5. A cursor drives every change with real clicks and drags. Drags are direct manipulation: while held, the value is computed from the cursor position; on release it springs back from wherever it was. The product's agent can be its own cursor (a gradient multiplayer cursor with a name pill, a short motion trail, lean with velocity, press scale + ripple on clicks) handing off from the user's cursor ("You").
6. The camera zooms so each state fills the frame. Empty frames are a bug; frame the content, then push, whip or pull back. Pull-backs reveal scale (one email → a grid of hundreds; one post → a wall).
7. Music-first timing. ~120 BPM, analyse the song with numpy for the beat grid, start on a downbeat, cut the song window so the drop lands on the product's big moment. Something happens on every beat; typing rides 32nds on a snare build; slams accelerate beats → 8ths → 16ths.
8. Every UI sound is synthesized and placed by its measured peak on the kick-peak grid.
9. Loops: when the piece loops, the last frame is the first frame — cursor position and speed included — or the loop stutters.

Look

• Palette: the brand's own design system (for Cheerful: warm near-white canvas, brand gradient #B04ADC → #E35C91 → #FF7247 → #FFB21E, lilac/peach light pools, ink #111827). Not "black plus one purple". One clean UI font.
• Banned (from the original brief, still binding unless Chan overrides): bouncy easing, particle bursts, glows, gradients on UI chrome, mismatched icon strokes, dead time, template looks.

Content

• Use real data. Chan: "You have access to the database. You can be creative and not just pick images." Pull real, defensible numbers read-only (creators reached, gifts sent, posts/views tracked) and show them as rolling odometers. Invent the micro-story (creator, brand, messages) but keep the big numbers true, and say which is which in the report.
• Customer logos one by one, boom boom boom — never a website-style "trusted by" row. One shape re-sized/re-coloured on every hit, only brands that really ran campaigns.
• Licensed media only. Stock video/stills from free-licence sources, checked per item (Mixkit has Restricted items mixed in). Never the product's own marketing images as a shortcut.
• Keep what he liked: for Cheerful the headline intro, the typed brief, the discovery search wall and the wordmark ending tested well — keep strong beats between versions and iterate the rest.

Process

1. Ask for inputs first (unless already given or he says "you decide"): 8–12 states the shape should become, palette (B/W, one accent, or the brand system), and a royalty-free ~120 BPM song. Offer "Director's cut / Director picks" options — he usually picks those.
2. Show the state list on the beat grid before writing code (bar/beat table: what happens, which element morphs into what, which sound).
3. Build one HTML file where every style is computed from time inside seek(t) — no CSS transitions, no timers, no state between frames.
4. Render one frame per beat before the full render, contact-sheet it, and fix anything off the grid, cramped, empty, hard to read or overlapping. Then check transitions at ±0.1 s.
5. Full render with Playwright subframes → ffmpeg tmix motion blur → 60 fps H.264 + mix.
6. Verify the MP4 itself (1-per-second storyboard, frame-difference scan for seams/glitches), serve it over the tailnet, and report: link, concept, scene list, real-vs-invented data, DECISIONs, caveats (audio is unheard — say so; logo/likeness permissions).

Gotchas that cost hours

• Never put will-change on anything the camera scales — text renders blurry.
• Springs start at their event: a target that must arrive at time T is issued one settle-lead early.
• hide() that zeroes opacity must be undone by the draw function, or elements stay invisible.
• Proportional digits jump in odometers: size each slot to its final glyph, clip only vertically.
• On the VPS: disk is tight (render in 4 s chunks and delete PNGs), Chromium is the first thing earlyoom kills (auto-retry, relaunch, hard-exit when done), a stray /tmp/inspect.py shadows the stdlib (run Python from the project dir), PreToolUse hooks block some ffmpeg glob strings.

Output

Deliverables in a deliver/ folder served with serve <dir> (tailnet URL, never localhost): the MP4, a 1-per-second storyboard, key stills, and the previous version for comparison.