Duration —1080×1080 · 1:1 squareHTMLSingle-file HTML + GSAP (seek-driven)
Run log
ModelClaude (agent build)
Renderer / stackSingle-file HTML + GSAP (seek-driven)
Iterations to final1
Render time34s (final render)
Format1080×1080 · 1:1 square
Published2026-10-25
What broke (and the fix)
Nothing broke — the first render shipped. Percentages land exactly at 42/28/19/11 (summing to 100), the WINNER tag sits clear of the other rows, and the 30%/70%/final stills are clean.
Original promptTested
Survey results — HTML prompt (poll data as a film)
Banned list —Beat grid —Build method —Self-review —
<direction>
An 8-second survey results film in single-file HTML (1080x1080 square, 30fps) answering WHAT SHOULD WE BUILD NEXT?. House style: near-black #0a0a0a, white type, ONE orange accent #ff5c1a, Arial/Helvetica bold, no webfonts. Four options (DARK MODE / MOBILE APP / TEMPLATES / API ACCESS) grow as horizontal bars to their vote shares; the winner takes the accent and a WINNER tag. Total: 1,204 SAMPLE VOTES. Sample poll, labelled SAMPLE on screen.
Banned: real poll data, fake voter counts presented as real, glows, gradients, pie charts, more than four options.
</direction>
<structure>
0.0-0.8s: question "WHAT SHOULD WE BUILD NEXT?" + "SAMPLE POLL" kicker in. 0.8-5.2s: four bars grow left to right in rank order, percentages counting up beside them. 5.2-6.4s: WINNER tag lands on the leading bar. 6.4-8.0s: footer "1,204 SAMPLE VOTES" + "SAMPLE POLL — NOT A REAL SURVEY"; hold.
</structure>
<build>
One single-file HTML composition rendered via HyperFrames, 1080x1080, seek-driven GSAP timeline, paused. Bars grow by scaleX from transform-origin left; percentages are object tweens. Render at 30fps.
</build>
<gotchas>
- Percentages must sum to 100 in the data, or the film fails its own arithmetic.
- Bars share one scale: the track width is 100%, so 42% fills 42% of the track — never rescale the leader to fill the frame.
- The winner tag must sit inside its bar's row, not float over the next option.
</gotchas>
Before you finish: step through the timeline at five points and fix anything that overlaps, clips the frame, or drifts off the beat grid.
A poll film must pass its own arithmetic: 42 + 28 + 19 + 11 = 100, and the bars share one scale (track = 100%). The winner tag sits in its own row, never over the next option.