Why did an AI-made product video pass every automated check and still look crammed, and what should the critic check instead?

Hypothesis

Our /brag skill turns a project into a 15 to 30 second launch video. Each draft is checked by critics that read frames and flag problems. The belief was that a video which passes those critics is ready to show. It was not.

What we built

• The video. A 32 second launch video for the agenda page of a community events site, about 3.9 MB with audio. The page pins its hero in a scroll scene, fades the hero out, fades the video in and loops it muted. A "watch the full video" button restarts it with sound.
• The failure. Versions 1 to 4 passed the critics. Robin's note on the result: box text was crammed towards the edges, button labels were not centred top to bottom, and nobody was tracking how fonts and boxes sat on screen. The critics checked clipping and overlap, not alignment, padding or type scale.
• A frozen layout checklist, now part of the skill and of every critic brief:
  • one shared grid: headline, caption and panel keep the same x position, width and top edge in every scene, so nothing jumps between cuts
  • card padding of at least 22 px top and bottom and 28 px left and right at 1080p
  • pill and button labels centred both ways
  • at least 16 px between rows
  • at most three text sizes, none under about 36 px at 1080p
  • critics compare neighbouring scenes for drift, not only clipping
• A 4 fps frame check. Pull frames at 4 fps into contact sheets with ffmpeg and look at every sheet, because short glitches hide between scene stills.
• Silent test browsers. No preview run may play sound on the user's machine.

Learnings

• A critic only finds what its checklist names. Ours listed clipping and overlap, so those were the only problems it could report. Missing items stay invisible however many rounds you run.
• Neatness can be written as a list of numbers. Padding, gaps, centring and type sizes all have values a critic can check. We have not yet measured how many problems the new checklist catches.
• Critics do not converge. More rounds do not give more agreement. We now freeze the checklist and stop after three rounds.
• Motion must not cover text. Make room first, then drop the card in. A card should never slide over another card's text.
• Pitch before mechanics. The code shows the buttons, the public page shows the promise. The skill now says to read the public page first and lead with its outcome.
• Framer Motion bug. On a sticky scroll scene, Framer Motion 12.43 ran a range-form opacity transform on the native scroll timeline, so the fade was wrong on the sticky scene. Using the function form of transform instead of the range form fixed it. We did not dig into why.
• The skill is the right place for the lesson. The checklist lives in the skill, so every later video starts with it instead of rediscovering it.

Log

• 2026-10-26: Video embedded on the agenda page. Preview deployed. Open question: the scroll preview is muted on purpose, because browsers block autoplay with sound, so a small unmute button is a possible next step.
• 2026-10-25: Layout checklist added to the brag skill after the "crammed text" feedback on the fourth version.