Skip to content

skill: decide video length from the content, with no 60-second default #30

Closed

Description

@umiyosh

Problem

SKILL.md sets the default length to "about 60 seconds (6-9 scenes)". Nothing in the code limits the length, but in practice every video comes out around a minute, and a theme that needs more is cut to fit. The figure came from the Claude Motion era; the html path has no reason for it, and even with Motion a longer video should be possible.

Proposal

  • SKILL.md: decide the length from the theme's content, with no upper limit. One mechanism: about 60 s; a theme with several steps or comparisons: a few minutes. A length the user gives always wins
  • When a video will be longer than about 2 minutes, tell the user up front how long rendering takes (about 5 min per minute of video today, see perf(build): capture html film frames on several pages in parallel #25), without stopping to ask
  • README (en / ja): say there is no upper limit and give the render time per minute

Check with a real video

Make one video of about 3 minutes end to end and confirm:

  • BGM: Lyria is asked for the film length + 10 s. Whether it returns a long enough track is unverified; if it is short, the mix loops it, and the seam may be audible
  • The html film: one page with many scenes. Watch for layout or quality problems as the page grows
  • Render time against the baseline in perf(build): capture html film frames on several pages in parallel #25

Part of #16.

Activity

  1. umiyosh commented on Oct 25, 2026

    @umiyosh
    OwnerAuthor

    Check with a 3-minute video

    Made "loop engineering" end to end on the html path: 20 scenes, 194 s, 73 sound effects, built with the parallel capture of #31.

    what result
    BGM Lyria was asked for 204 s and returned 169 s, then 178 s on a retry. A film longer than about 3 minutes gets a BGM that loops; the restart is at 2:58 in this film. Not yet judged by ear
    html film One page with 20 scenes worked. check found 6 issues (nodes under a ring drawn on the wrong layer, an intended strike-through), all fixed; stills showed one more (a moving dot resting on a label), fixed
    render time 324 s for 5,820 frames with 7 pages (load average about 10 from other jobs). The same length at master's speed would be about 15 min
    layout check 5-8 s for the whole 194 s film with 7 pages
    sound effects all 73 peak at least 5 dB over narration + BGM in their window (median 9.4 dB)

    Follow-up for long films: the BGM. Options are asking Lyria for two parts and crossfading them, or crossfading the loop point instead of restarting hard.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions