Watermelon rubber band burst simulation
- Views
- 426.1K
- Likes
- 104
- Replies
- 16
- Saves
- 32
An interactive 3D browser toy where tapping a watermelon wraps rubber bands around it until it bursts, built with Claude Sonnet 5.5 Extra from a single detailed prompt, with an estimated API cost of $4.96.
Full prompt5,567 chars
Create "The Watermelon Test": a single-file interactive 3D toy in HTML. A striped watermelon sits on a warm beige floor; each tap wraps another rubber band around it, the fruit squeezes into an hourglass, and at some point it bursts. Deliver the full working HTML, not a mockup or code fragment. STAGE AND LOOK - WebGL via Three.js (bundled inline or pinned from a CDN). No image or audio files. Google Fonts are the only external resource. - Warm studio look: page background is a soft radial gradient from #f4f0e8 (centre) to #e6dfd3 (edges). Transparent canvas over it. Floor is a shadow-catcher only, plus a soft contact-shadow blob under the fruit. One key light from above-right with soft PCF shadows, a warm fill light, a hemisphere light, and a subtle room environment for reflections. Neutral tone mapping so the greens stay clean. - Camera pitched about 23 degrees down, 28 degree FOV. Size the scene so the fruit is about 55% of the width on a portrait phone and about 40% of the height on desktop. TYPOGRAPHY - Title "THE WATERMELON TEST" in an extra-wide bold grotesque, all caps, tracked (Archivo at 125% width). Subtitle "How much pressure can it take?" in a light italic serif (Newsreader). Small mono labels (JetBrains Mono, wide letter-spacing). Big italic display serif for the end screen (Instrument Serif). Ink #1d1a16, accent red #d2404c. Top-left title, top-right RESET and SOUND buttons with small line icons, bottom-centre HUD. THE WATERMELON - Egg-shaped surface of revolution about 2 units long, max radius 0.74, right end slightly narrower. Rebuild its vertices and normals on the CPU every frame from a radius profile, so it can pinch. - Procedural rind shader driven by each vertex's rest position, so stripes stay glued to the fruit while it deforms: about 15 wavy dark-green stripes (#265c33 range) on pale, slightly mottled green (#aed087 range), soft stripe edges, fine grain as a bump, stripes converging into a dark pole at each end, waxy clearcoat. - On load it drops from above with a bounce and a squash, then rests. Dragging rotates it freely with inertia (trackball style). Mouse and touch. RUBBER BANDS - Tap on the fruit adds one band at the tapped position along its long axis; taps elsewhere do nothing. Ignore taps that were drags. Also Space or Enter adds a band. Hover cursor changes over the fruit. Each tap shows a short expanding ring at the pointer (tan outline). - Bands are thin flat ribbons (width about 0.04 to 0.056) that hug the deformed surface, each tilted 1.5 to 12 degrees in a random direction so they cross. A new band starts about 1.5x too large and snaps onto the fruit with a spring overshoot. - Bands slide along the surface and stack into a bundle around their centroid. Bundle width about (0.22 + 0.0058 x N) in rest units. Each band sits at a slightly different radial layer. - Colour goes from tan (#d9a66a) toward cream (#f6e9cc) as pressure rises, strongest in the middle of the bundle, with random variation so you see both. - Squeeze depth D = 0.60 x (N/65)^1.15 pinches the fruit around the bundle (waist about half the end diameter near 53 bands), with the ends bulging slightly. Every tap adds a small spring kick. Above about 40 bands add a faint tremble. HUD - "BANDS:" with the count (red once above 0), a caption that changes with a quick fade, and a thin progress line that fills N/65 and turns red at 40 or more. - Captions: 0 "Tap the watermelon to add pressure", 1+ "Everything seems fine.", 13+ "Probably still fine.", 28+ "Bro is under pressure.", 40+ "This feels irresponsible.", 49+ "Maybe stop." BURST - Each round picks a random burst count between 50 and 62. On the last band: 0.34 s of violent tension (shake, bands whiten), then it bursts. Hide the fruit, shake the camera, and spawn rigid-body debris: - 7 jagged curved rind shells (striped green outside, pale green-to-white edge, flesh inside fading from white at the rim to coral pink), - 15 faceted low-poly red flesh chunks of varied size, flat shaded, - about 26 flat glossy black seeds, - the rubber bands as curly strands and loose loops, - about 130 small pink droplets that turn into flat translucent pink splat decals on the floor when they land. - Debris physics: gravity, contact points against the floor with impulse and friction, restitution, damping, sleeping when still. Keep pieces inside the visible floor area, with some cropped at the screen edges. - After about 1.9 s fade in a centred overlay: huge italic "Well…", a small mono line "53 BANDS." with the number in red, and a black "DO IT AGAIN" button. Hide the HUD. RESET and DO IT AGAIN start a new round with a new random threshold. SOUND (procedural Web Audio, no files) - Create the audio context only after the first user gesture. SOUND button toggles it (on by default), icon shows muted state. - Band snap: a short filtered noise click, a triangle "twang" dropping in pitch (pitch rises with pressure), and a dull low thump; small random variation each time. - A slow, quiet creak now and then once past about 30 bands. - Burst: sub thump, noise crack, wet splats, and a flurry of band snaps. Soft thuds when debris lands, rate limited. - Limit simultaneous voices and disconnect every node when it ends. QUALITY - Works on desktop and phone (portrait and landscape), pixel ratio capped at 2, no horizontal scroll, side gutters at least 16px. Skip geometry rebuilds when nothing is moving. Fallback message if WebGL is unavailable. Visible focus states and aria labels on the buttons and canvas. Real content only, no placeholders.
Source: the creator's reply under the post. Kept in its original language, unedited.
Want more prompts like this one?
Get the 50 most-viewed Claude 5.5 video prompts on one page: the full original text, the creator and the post it came from. We'll only email you when there's a new prompt worth trying. Unsubscribe anytime.
Tools the creator mentioned
Original post: https://x.com/vib3coded/status/2104928244382277727
All works are open data at GitHub · awesome-opus-5.5-video.
Model attribution is as stated by the creator and was not independently reproduced. View counts are a snapshot from 2026-11-19. The work belongs to its creator; DM @GoSailGlobal on X for corrections or removal.
More 3d worlds
0:17AI-generated running cow animation
1.4M views
0:18Jelly watermelon slicing simulation
634.7K views
1:00Playable boat scene through Japanese landscapes
489.1K views
0:15Visualization of humanity's biggest scam
445.8K views
1:31Pelican riding a bike in Three.js
421K views
0:33Floor plan to 3D interior design tool
373.9K views
Related articles
A Practical Guide to Claude Code Skills: Your Secret Weapon for Developer Productivity
A deep dive into Claude Code's Skills system, covering Design Skills evaluations, SEO audits, code quality guardrails, and other hands-on techniques.
Anthropic Claude Certified Architect Complete Roadmap: 8 Free Courses + $99 Exam
Anthropic launched the Claude Certified Architect certification on March 12, 2026, with 13 self-paced courses — all completely free, no registration required, no paywalls. The exam itself costs just $99. This post walks through a 12-week roadmap covering the 8 key courses in Weeks 1–9, with direct links to each.
Same Name, Different Game: Comparing Two SEO Audit Skills and Which to Use
The Claude Code Skills ecosystem is growing fast — fast enough that naming collisions are starting to happen. Two skills both called seo-audit exist, and they take completely different approaches. Here's a side-by-side breakdown and recommendation.
Meta's Official PyTorch Tutorials: The De Facto 'Deep Learning Field Manual' — 9 Steps from Training to Deployment
Meta never built an AI academy, but it turned the official PyTorch tutorials into an industry standard — the de facto deep learning field manual. Completing them is the equivalent of working through a full modern deep learning curriculum. This post maps out the recommended 9-step path plus Llama companion resources.