A typed prompt becomes a moving image gallery

A tiny request on an open field yields a moving result, then a gallery demonstrates range. A follow-up changes an existing example; the image payoff grows into a brief close-up and a restrained brand finish. Picture changes and card choreography carry the story.

Why it works

• The request is readable in the opening and the first result appears around 3 s, before a feature claim.
• Shot 1 can contain the full request, gallery and revision because reframes at frames 467 and 775 and a change at 540 refresh its white stage.
• The short full-frame insert at shot 2 creates a scale shock before shot 3 returns to a single card and typography.

Tools you'll need: an image model (GPT Image, Midjourney) · Remotion for motion graphics · a licensed or generated music track · Claude alone: partly

Hook formula (inferred): typed experiment: Can {{PRODUCT}} make {{FIRST_REQUEST}}? Keep the request specific enough to verify from one picture. The opening's white space focuses the eye on changing letters.

Retention map (inferred): re-hooks at the follow-up in beat 3 and range escalation in beat 4; visual payoff in shot 2; text and mark in beat 5; clean stop.

Mix (inferred): almost all white UI/card choreography in shot 1 and shot 3; shot 2 is a compact full-frame detail. Rhythm (inferred): motion within cards and horizontal staging supply changes during long shot 1. The two measured hard boundaries cluster late; the end card changes remain inside shot 3.

Rules this editor follows (inferred):

• Let each new card demonstrate a different possibility, then return to one subject before the revision.
• Keep the white stage and use scale changes to distinguish a result, an array and the final proof.
• Reserve the full-frame colour fill for shot 2, then restore open space.

• Look (inferred): quiet high-key field, soft raised controls, rounded portrait cards; image content supplies saturated accents. The reference's predominantly near-white field and blue/pink accents are evidence of roles, not brand values to reuse.
• Graphics and overlays (inferred): request pill with small controls, consistent card corner radius, restrained shadows; moving content stays within masks.
• Transitions (inferred): preserve the measured soft three-frame change, instant reframes and hard internal changes; keep the two measured shot boundaries as cuts.
• Branding and end card (inferred): shot 3 replaces the performer with a short line, then {{FEATURE}} and {{LOGO}}; finish with {{LOGO}} and {{CTA_URL}} as text on {{BRAND_BG}}. No fade to black.

      0s         10s        20s        30s
mean  ███▇▇▇▇▇▇▇ ▇▆▇▇▇████▇ ▇▇▇▇▇▇▇▅▆▇ █████
p95   ██████████ ██████████ ███████▇██ █████

:root { --brand-bg: {{BRAND_BG}}; --brand-accent: {{BRAND_ACCENTS}}; --brand-ink: color-mix(in srgb, var(--brand-bg), black 92%); }
.stage { background: var(--brand-bg); color: var(--brand-ink); }
.request { border: 1px solid color-mix(in srgb, var(--brand-ink), transparent 92%); border-radius: 999px; background: color-mix(in srgb, var(--brand-bg), white 10%); box-shadow: 0 9px 30px rgba(0,0,0,.08); backdrop-filter: blur(8px); }
.card { overflow: hidden; border-radius: 18px; box-shadow: 0 12px 36px rgba(0,0,0,.12); transform-style: preserve-3d; }
.card img, .card video { width: 100%; height: 100%; object-fit: cover; }
.hero-glow { filter: drop-shadow(0 0 24px color-mix(in srgb, var(--brand-accent), transparent 45%)); }

Measured: plate 8 reports five words over 15%, while v1's transcript consists of punctuation and its pitch estimates are unreliable. No legible narration is needed to understand the picture. Voice direction (inferred): no voice track in the new film. Let request text, results and music carry it; if a product requires accessibility narration, treat that as a separate version and time captions to speech.

• Music (inferred): playful tonal bed with clipped percussion and airy synth notes. The tracked pulse is steady enough for composition, yet measured cuts do not follow it.
• Sound effects (inferred): light key clicks for the opening, soft swishes for card movement, firmer impact near the revised result, and small punctuation for the late text changes. Do not duplicate events marked voice? as effects.
• Silence and space (inferred): leave space around the instant close-up and the isolated mark; keep the bed low enough for readable UI pacing.

▇▆████████ ██████▇▇██ ██████████ ████▅▁

• Title (inferred): “Introducing {{FEATURE}} for {{PRODUCT}}” — plain launch naming.
• Cover / thumbnail (inferred): one original result card against {{BRAND_BG}}, a compact {{PRODUCT}} name, and one high-contrast accent; avoid the reference's subjects and candy lettering.
• Post copy (inferred): a one-line change claim, one concrete capability, then {{CTA_URL}}.
• Audience signals: the supplied capture has 135,117 views; no replay heatmap. Comments in v1 react to the usefulness of short moving images and to specific playful examples.

Picture-led: there is no spoken word budget. Give each request enough on-screen time to read at phone size; keep the end line short.

[beat 1 · shot 1 · ≤8 readable words] {{FIRST_REQUEST}}
[beat 2 · shot 1 · no new claim] show distinct original results
[beat 3 · shot 1 · ≤8 readable words] {{FOLLOWUP_REQUEST}}
[beat 4 · shots 1–3 · no feature list] show the change and its most expressive result
[beat 5 · shot 3 · ≤4 words] {{FEATURE}}
[beat 5 · shot 3 · final] {{LOGO}}  {{CTA_URL}}

Two routes (inferred)

Minimum: Claude alone in Remotion. Use one code stack for the pill, gallery, CSS depth, animation, end card and a synthesised original music bed. Supply real {{PRODUCT}} captures and original {{RESULT_ASSETS}} as inputs; Claude alone cannot make those truthful assets. Static cards with subtle CSS motion can reach roughly 80% of the composition but give up expressive subject motion and polished image variation.

Full: with the toolkit. First create a consistent set of original image results with an image model, then obtain or animate short loops from those assets. Export 4:5 or portrait RGBA PNG sequences at 30 fps, at least 1080 px tall. Pass asset IDs and camera/card keyframes as JSON to Remotion; composite UI, card masks, gallery, text and logo at 1920×1080, 30 fps. Generate or license an original music WAV at 48 kHz plus a click-track WAV aligned to the timeline cue frames; mix and render H.264 MP4. The click track is a guide, not part of the master.

Asset honesty: use real captures of {{PRODUCT}} or its UI rebuilt in code; if a result does not exist, label a type tile clearly. Never generate product UI, screenshots, metrics, reviews or outcomes as false evidence. Here, keep the before and after subject recognisably the same asset lineage.

Script prompt

For {{PRODUCT}} and {{FEATURE}}, write {{FIRST_REQUEST}} as a concrete request of at most eight words. Write {{FOLLOWUP_REQUEST}} as a visible change to one existing result, also at most eight words. List original {{RESULT_ASSETS}} that demonstrate different outcomes without using the reference's subjects, copy, UI, logo, palette or music. Keep the on-screen ending to a short feature line, {{LOGO}} and {{CTA_URL}}.

Image prompt (inferred): “Make each item in {{RESULT_ASSETS}} one subject per portrait frame, consistent framing and identifiable silhouettes, no letters or interface. Make the revised result the same subject with one clearly visible change. Use {{BRAND_ACCENTS}} by role, with distinct hues across adjacent cards.”

Music prompt (inferred): use plate 8's original bed prompt; deliver an instrumental WAV and separate cue stems.

How it's built (inferred)

Signature move: the answer expands into a travelling gallery. In shot 1, keep the canvas fixed and move a parent card track behind a 1920×1080 white stage. The detected shot-level −217.6% x travel in plate 5 is a composite of multiple layouts, so keyframe individual cards instead of applying that as one continuous camera pan. Use the measured opening: hold the pill nearly static through frame 64, pull it to ×0.90 over frames 64–72, then place the first card below it. For the first result, enlarge within its rounded mask, reduce it again around frames 120–135, and introduce neighbouring cards from the sides. For lateral travel, use cubic-bezier(.42,0,.25,1), with position and scale keys at each card arrival. Keep image content moving on ones where practical; the film's measured moving cadence averages 22.9 unique frames/s, so hold selected frames for two rather than smoothing every loop at 60 fps. Add directional motion blur on the card track only, roughly 8–14 px at peak travel, never on readable text.

Other moves and hand-offs. Frames 413–415: blend the paired pictures across 3 frames inside their masks; the strips show a gradual change. Frame 467: instant reframe to the larger existing card, with no invented tween; use the plate 5 jump as a reset of card positions. Frame 540: switch to the revised result on that frame, with the old and new assets aligned at their subject centre. Through the hero, animate the card crop and accent within its mask, then spread cards again for beat 4. Frame 775: instant ×0.90 reframe of the array and reveal the deeper layout behind it. Frames 822 and 844: the only measured shot-boundary hard cuts, entering the close-up and then the performer card. In shot 3, the element-motion table's ↓ travel is an aggregate across artwork and end graphics; the sheet shows the performer card descending out before the end line. Build it with a parent y translation and mask, easing cubic-bezier(.55,0,1,.45) from about frame 868 through 892. At frame 893, hard switch to the short line; at frame 933, hard switch to the feature lockup. Later, scale the mark into its isolated ending around the plate 5 segment at 33.33–34.13 s, then hold through the final frame. The sheet suggests a later change to the standalone mark; the Timing facts list no cut after frame 933, so keep it as an element replacement within shot 3. The measured shot 3 camera is low confidence: leave the white stage locked and animate the contained card and type.

Gotchas. Keep the same subject between old and revised cards; otherwise the follow-up reads as another random gallery item. Pre-render loops with shared crop coordinates and alpha before animation. Use static font files for any outlined or gradient feature letters to avoid seams, and subframe sampling for fast card swishes at 30 fps. The bright white stage must remain stable through the internal cuts; the darker close-up is confined to shot 2 and the measured brightness step.

Build order

1. Map beats 1–5 to the measured shot frames and the six internal change frames in plates 4–5.
2. Make and approve {{RESULT_ASSETS}}, including one before/after pair and one full-frame detail.
3. Build shot 1's pill, card component, mask and travelling gallery; add its two request text builds.
4. Build shot 2's close-up and shot 3's performer-to-type hand-off.
5. Place original bed and cue accents using plate 8, then mix to the delivery target.
6. Render and run plate 14 checks at phone size.

• Shot-boundary cuts at frames 822 and 844, within ±1 frame; internal changes at 413 (3 f), 467, 540, 775, 893 and 933.
• Request and end text cap height ≥38 px in the 1920×1080 deliverable; other aspects meet plate 15 minima.
• Shot 1 shows an identifiable same-subject revision after frame 540; the card masks have no exposed edges during travel.
• Shot 2 lasts 22 frames and its darker field does not persist into shot 3.
• Cue accents land within ±2 frames of chosen plate 8 events; the two measured shot cuts are not forced onto beats.
• YouTube master at −14 LUFS integrated and true peak ≤−1 dBTP; no clipped key clicks.
• The final image stays bright; no fade to black.
• All supplied result assets are original or licensed, and all product captures are genuine.

Brand slots

What this film avoids (inferred): no presenter, long voiceover, feature list, subtitles, stock footage, QR code or store badges. Its proof is shown as image results.

Brand-distance warning (inferred): a near-white stage plus the reference's exact input chrome, animal examples, pink effect and candy-coloured feature letters would read as the original brand. Change the UI silhouette and use your own image subjects.

Adapt the aspect (inferred)

• 9:16: re-lay out the gallery vertically because a centre crop keeps only the middle 31.6% width and one measured text box. Stack one dominant card over a smaller preview; move requests into the middle 80% width. Use ≥22 px cap and keep essential text away from top/bottom platform controls and the right-side action rail.
• 1:1: centre crop the stage, then move the lost feature lockup fully inside the square and reduce card spread. Use ≥22 px cap; keep the CTA away from lower feed overlays.

Adapt the length (inferred)

• 15 s: keep beat 1's request/answer, beat 3's same-subject revision, shot 2's insert and beat 5's finish; compress or omit the extra range examples in beats 2 and 4. Preserve the signature gallery movement in a compact form.
• 30 s: keep all five beats; reduce dwell on repeated gallery examples, preserving the follow-up, close-up and final mark.
• 60 s: add new original examples and one more clear revision inside beats 2–4; do not stretch the typing or final brand hold simply to fill time.

The creator's signature (don't copy): reference UI, particular animal results, exact phrase, logo, glossy lettering, colour combination and music. The transferable format (do copy): typed test → visible result → moving range → same-subject revision → short detail insert → spare brand finish.

Source: https://www.youtube.com/watch?v=bJldMYfRvMQ · measured with dissect 1.1 and timeline v2 on 2026-11-09 · (measured) = read off the film · (inferred) = construction reading