---
title: "The Prompt Behind Triptych: A Scroll-Scrubbed Altarpiece Site Made With Claude and Higgsfield"
description: "A copy-pasteable prompt for a pear.no-style cinematic site: eight Higgsfield clips chained into one scroll, a seam that opens the page, and hairlines locked to the painting."
date: "2026-11-22"
type: "prompt"
tags: ["higgsfield", "claude-code", "webgl", "scroll", "seedance", "motion", "hero"]
output: "/lab/triptych"
ogImage: "/lab/triptych/og.jpg"
---

Triptych is a studio that doesn't exist, with a site that opens like an altarpiece. You scroll and a donor's hand lifts off a closed panel, the wings swing open on their hinges, the camera pushes through the gilded frame into the painted garden, flies past the fountain into the sun, and comes out on a candlelit desk where a signet presses a wax seal. Then the wings close again around the contact form. You can [scroll it at avivashishta.com/lab/triptych](/lab/triptych).

It's the sequel to [the pear.no-style site I built on Replit](/blog/prompt-scroll-scrubbed-cinematic-site-webgl-replit). Same engine: one sticky stage, one scroll number, WebP frame sequences in WebGL, real HTML text. This time I built it in Claude Code with the Higgsfield MCP, so Claude could generate the stills, animate them, check every clip frame by frame and cut them into sequences without me leaving the terminal. pear.no is still the reference, and the thing I went after was what makes it feel expensive: the borders, the edges and the way the type arrives.

<div class="highlight"><p><strong>The short version:</strong> chain your clips with shared keyframes, so the end frame of one clip is the start frame of the next and the whole site is one camera move. Draft every clip at 480p and only pay for 1080p once the motion holds. Then make the hairlines part of the picture: track one line inside the footage frame by frame and draw the page's borders from it, so the layout and the film are the same object.</p></div>

## The prompt

Paste this into Claude Code with the Higgsfield MCP connected (or the CLI installed). It's long because every clause is something that went wrong without it.

<div class="promptbox">
  <button class="copybtn" type="button" data-copy>Copy prompt</button>
  <div class="prompt">
<p>Build a single-page, scroll-scrubbed cinematic site in the class of pear.no for a made-up studio called Triptych. The theme is a 15th-century Flemish altarpiece that opens. Generate every image and clip with Higgsfield. Vanilla JS, hand-written WebGL1, native scroll, no GSAP, no three.js, no Lenis, no build step. Show me the credit cost of each Higgsfield batch before you spend it.</p>
<p>THE STORY (one continuous camera move, five 16:9 clips plus three 3:4 panel clips): S1 a donor's hand rests on the glowing gold seam of a closed altarpiece, withdraws, and the camera pulls back to the whole altarpiece in a candlelit chapel. S2 camera locked, the wings swing open on their hinges and light floods out of a painted paradise garden. S3 the camera dollies through the gilded frame into the garden until the painting fills the screen. S4 it flies past the fountain, over the hills and through the clouds into the sun until the screen is gold. S5 on a candlelit desk, a hand presses a gold signet into vermilion wax on a letter and lifts it to reveal the studio's arch mark. P1 to P3 three workshop panels: a gilder laying gold leaf (identity), a glassmaker holding stained glass to the light (film), masons raising a cathedral tower (sites).</p>
<p>KEYFRAMES (GPT Image 2.6, high, 2K): one shared style suffix on every prompt: "Early Netherlandish oil painting realism in the manner of a 15th-century Flemish master, luminous oil glazes, fine panel craquelure, lapis blue, vermilion, verdigris green and burnished gold leaf, soft directional light, no text, no lettering, no watermark". Make the closed altarpiece first. Make every other keyframe that shows the same object WITH that image as a reference (the close-up, the open altarpiece, the inside of the painting), so the paintings on the wings never change. Keep the left third of every frame dark or calm for the headline. Never paint figures in the sky: winged figures got my clips flagged by the safety filter.</p>
<p>CLIPS (Seedance 2.5, omni_reference, 5s, no audio): use start_image AND end_image. The end frame of each clip is the start frame of the next, so the site is one take. One camera move per clip, said in plain words ("smooth slow dolly pull-back, one axis, even speed"). Generate every clip as a 480p draft first. For each draft, build a contact sheet and measure frame-to-frame difference with ffmpeg; reject any clip with a spike over about 2.5x the mean (a cut) or a cross-dissolve where two versions of the same painting overlap. Only then finalize at 1080p, resending the full original request (prompt, medias, duration, aspect ratio) with draft_job_id.</p>
<p>FRAMES: cut every clip into 121 WebP frames in two tiers (16:9: 1600 and 1024 wide; 3:4: 720 and 540 wide). Hard budget 12 MB per desktop sequence: lower quality before frame count. Track the gold seam in S1: in each frame find the narrowest bright warm vertical line, starting from the strongest peak in frame 1 and then searching only a few pixels around the previous position, and save its x per frame to seam.json.</p>
<p>ENGINE: one section 2100vh tall with a sticky 100vh stage. One requestAnimationFrame loop reads scrollY once, lerps a 0..1 progress at 0.12, and drives everything from that one number through a TIMELINE of ranges. Frame loader: halving-stride order, a priority window of 24 around the wanted frame, nearest-decoded fallback so nothing is ever black, at most 6 requests in flight, retry with backoff (never re-queue a failed frame immediately), range-gated by where each sequence is used, and release sequences more than 0.25 away. One WebGL1 fullscreen triangle with two textures and a mix: crossfade, a halftone dot dissolve, and a ring wipe with a gold rim. Always on: highlights pulled toward cream, vignette, dither, film grain, a warm candle glow following the cursor. Cover-fit with a focal x so phones crop toward the seam, not the centre.</p>
<p>HAIRLINES (a 2D canvas over the GL canvas, 1px, white at 16%): a top rule that stops at two thirds, a left rail, four-point stars wherever two lines cross. Draw the gold seam from seam.json through the cover-fit math so it sits exactly on the painted seam, extend it to the screen edges, split it in two as the wings open so each half rides a wing out to the frame, sweep a full-width rule through the garden, draw a pear-style grid around the three panels, drop column dividers for the terms, and at the end thread dashed gold lines from each form field to the seam.</p>
<p>TYPE: real HTML. Instrument Serif headlines rise line by line out of overflow-hidden masks with an SVG feTurbulence ink filter that settles to zero as they land; the hero rises word by word. A mono eyebrow on the rail swaps its label behind a retracting bar. Pills for kickers. The FAQ answers open with a tearing displacement filter.</p>
<p>INTRO: start on black. Draw one gold hairline exactly where the seam will be, with a pulsing star and a loading count, and wait for the first 24 frames. Then open the picture from that line outward in the shader, with a glowing gold edge, starting blurred and pushed in and settling sharp. The headline, rules and nav follow.</p>
<p>HOLD THE HERO OBJECT: after the seal lifts, hold the finished seal with a slow push in before any transition, so the payoff is on screen alone.</p>
<p>FALLBACKS: the default CSS is a normal readable page (one section per screen with a still behind it); the cinematic stage only switches on under html.js without reduced motion. ?static forces the fallback. WebGL lost falls back to 2D canvas crossfades.</p>
<p>Start with the closed altarpiece still and its cost. Show me each draft's contact sheet before you finalize it.</p>
  </div>
</div>

## Concept 1: the site is one camera move

pear.no feels like a film because you never see a cut. The Replit version had separate clips and hid the joins with shader transitions, and you could feel each one starting over.

The fix is in the keyframes, not the code. Seedance takes a start image and an end image, so I generated the stills first and handed them out like relay batons:

| Clip | Start frame | End frame |
| --- | --- | --- |
| S1 hand lifts, pull back | close-up of the seam | closed altarpiece |
| S2 wings open | closed altarpiece | open altarpiece |
| S3 into the painting | open altarpiece | inside the garden |
| S4 into the sun | inside the garden | a field of gold |

The closed altarpiece is the end of S1 and the start of S2, so frame 121 of one sequence is frame 1 of the next and the switch is invisible. Only two transitions are left on the whole page (gold into the desk, and the desk back to the altarpiece), and both are deliberate.

## Concept 2: if the start and end disagree, the model cross-dissolves

My first S1 had a ghost in it. About a second in, the angel and the Virgin on the wings faded into slightly different versions of themselves, because I'd generated the close-up and the wide shot from two separate prompts and the model painted the Annunciation twice. A video model asked to get from one to the other will happily dissolve between them, and when you scrub back and forth that reads as a glitch.

So nothing that shows the same object is ever generated from text twice. The wide shot came first; the close-up was made with the wide shot as its reference image ("same altarpiece, same paintings, recompose as a closer shot"); the open altarpiece and the inside of the garden were made the same way. After that, every clip ran without a morph.

Claude caught this by checking the clips the way a machine can: a contact sheet of the whole clip, plus the mean pixel difference between neighbouring frames. A cut shows up as one spike, and a cross-dissolve shows up in the contact sheet as two paintings at once.

## Concept 3: draft at 480p, pay for 1080p once

A Seedance draft is 480p and 15 credits. Finalizing it at 1080p is 80, and it costs the same 80 at 720p, so there's no reason to finalize lower. Every clip went through a draft first, and only clips that passed the contact sheet and the motion check got finalized.

One detail that cost me a round trip: finalizing needs the whole original request again (the prompt, the start and end images, the duration and the aspect ratio) along with the draft's id. The id alone fails validation. It doesn't charge you, but it's an easy thing to lose ten minutes to.

## Concept 4: the border is in the painting

This is the part I care about most. The altarpiece has a thin line of gold light where its two wings meet, so I tracked that line through every frame of S1 (the brightest narrow warm vertical, searched only near where it was a frame earlier, so it can't jump to a candle) and saved its x position to a JSON file.

The hairline canvas runs that number through the same cover-fit math as the shader, so a 1px gold line on the page sits exactly on the seam in the footage and extends it to the edges of the screen. When the wings open, the line splits and each half rides a wing out to the frame. At the end, the dashed lines from the form fields run to it, and the wordmark at the bottom is positioned so its own gold seam lands on the altarpiece's: "Trip" on one wing, "tych." on the other.

Everything else is pear's vocabulary: a top rule that stops short, a left rail with ticks, four-point stars wherever two lines cross, and a grid around the three panels with a star at every corner.

## Concept 5: an intro that opens

My first intro faded in from a blur, and it felt like a page loading. Now the page starts on black with one gold hairline drawn exactly where the seam will be, a pulsing star at its centre and a mono count while the first frames arrive. When they're in, the shader opens the picture from that line outward with a glowing gold edge, blurred and pushed in, settling sharp as it opens. The headline then rises word by word.

The site begins as the same line it's about, and it waits for real data rather than a timer, so it never opens onto half a frame.

## Concept 6: give the hero object a hold

In the first build the signet lifted and the halftone dissolve started on the very next frame, so nobody saw the seal. A clip that ends on its payoff needs the scroll to stop moving the film for a moment. Now the seal clip plays a little faster, then holds its last frame with a slow push in while the terms clear away, and only then dissolves. The same idea holds the gold field while the three panels rise out of it.

## Concept 7: load like a film, fail like a website

Eight sequences of 121 frames is too much to hold at once, so each sequence is registered against the stretch of scroll that uses it, loads coarse-first (every 32nd frame, then every 16th) when you get near it, and is released when you're far away. The one bug worth repeating: a missing frame was re-queued the moment it failed, so a single 404 became thousands of requests a second. Failed frames now retry with a backoff and go back in the queue only after a pause.

## What it cost

About 817 Higgsfield credits in total. The eight 1080p finals were 640 of that (80 each). The rest was drafts at 15, high-quality 2K stills at 2 each, and a first round of cheap 1K trial frames for three candidate themes (an altarpiece, kintsugi and an orrery) before I picked one. The page is about 77 MB of WebP frames on desktop and 41 MB on phones, but a visitor only downloads the stretch they're scrolling through.
