Skip to content

docs(rich): have the critic read a light storyboard before the film is written - #89

Merged
umiyosh merged 1 commit into
masterfrom
docs/rich-storyboard-critic
Nov 17, 2026
Merged

umiyosh merged 1 commit into
masterfrom
docs/rich-storyboard-critic

Conversation

@umiyosh

@umiyosh umiyosh commented Nov 17, 2026

Copy link
Copy Markdown
Owner

Summary

/motion-render:rich now has a critic read a light storyboard before the film is written.

The final critic pass on the contact sheet stays. Its job is now drawing and motion. What a scene shows is already settled before the film is written.

Part of #87.

Why

In the benchmark run in #87, 7:13 of the 48 minutes went to applying the critic's top 3 and verifying again. All three findings were about what the scenes show, and they could have been caught before the film existed. The guide asked for a plan.md, but nothing reviewed it, and that run skipped it.

Experiments in #87

  • Same findings from text. A critic given only a text storyboard of that run's draft named the same three problems in both runs.
  • Fixed problems are told apart. On the fixed version, it named only the problem that had been left in.
  • A light storyboard is enough. A full plan took about 16 min to write. A storyboard with only the five per-scene fields took 4 min, and its critique was still useful.

Changes

File Change
skills/rich/SKILL.md Step 3 now writes two files: plan.md (the intent, a few lines, never shown to the critic) and storyboard.md (only the five per-scene fields). Step 4 starts the storyboard critic in the background with bgm. Step 5 reads the kit and arranges the score while the critic reads, fixes the storyboard's worst 3, then writes the film. Step 7's critic now looks at drawing and motion.
references/html-film.md "見直し" becomes two reviews. (1) The storyboard review, before the film, with five questions; the critic gets storyboard.md, the narration and the scene times, not plan.md. (2) The final review on sheet.png, unchanged, plus a note to check that a finding is not only an artifact of the sampled frame times. "絵コンテと秒の計画" and "場面を分けて書く" now point at storyboard.md.
references/pipeline.md The step table adds the storyboard review, the storyboard fixes and the final review, and says why the critic does not add waiting.

Measured

One trial run (2026-11-18) followed this procedure. It made a 66.9 s film at 124 BPM, with one stop, two drops, 5 sound effects and a 42.0 KB film page. Last time the page was 42.7 KB.

Measure Benchmark run Trial run
Request to MP4, without environment work 42:00 32:17
Writing the film 10:05 6:07 (+1:41 self-review edits)
Rework from the critic's findings 7:13, scenes rewritten 1:18, storyboard text only
Waiting on a critic with nothing to do 2:56 0:47
Final critic's top 3 about what a scene shows 3 of 3 1 of 3

motion-render times for the trial: total 28 min 55 s from the first command, Claude 25 min 20 s, waiting on commands 3 min 34 s. The details are in the trial comment on #87.

The comparison has two caveats:

  • It is one run, on a different topic.
  • The same session had used the kit the day before.

Test plan

  • make check
  • One rich film made by following the changed guide; layout: ok, sync: ok
  • Next rich film: confirm the storyboard review runs in the background and the final critic's top 3 are mostly about drawing

@umiyosh
umiyosh merged commit fe9986a into master Nov 17, 2026
1 check passed
@umiyosh
umiyosh deleted the docs/rich-storyboard-critic branch November 17, 2026 17:07
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant