Learn style

Turn a rotoscoped clip into a style pack: the numbers that make the style what it is, and a check
that tells whether a new film shares them. The pack holds statistics only; never copy traced
geometry, source frames or the clip into it. Read
`reference/statistics.md` first: what each statistic measures, how a
band is set, which statistics separate the woodcut style from a filtered near miss and from other
styles, and what the check still cannot separate.

Requirements: the plugin README's (run the scripts the way it says; /animation:setup checks
them) and a rotoscope work directory (d/dNNN.json traces and
src/dNNN.png drawings). The scripts are ${CLAUDE_PLUGIN_ROOT}/skills/learn-style/scripts/learn.py,
${CLAUDE_PLUGIN_ROOT}/skills/learn-style/scripts/controls.py and the shared
${CLAUDE_PLUGIN_ROOT}/scripts/inkstats.py.

Pack format

styles/<name>/ in the plugin, tool-neutral:

File	Written by	Holds
style.json	learn.py, then you	palette (ink, paper, tone ramp), knobs (the seven-knob schema), stats (p10/p50/p90), timing, ref (the source's values), bands, check, measured_only, heldout (how each band was set and how it tested), brush, credit, rotoscope_fits
STYLE.md	you	the style in words for a reader: credit, knob table with each value marked measured or judgment and pointing at its style.json key, how to author in it, the validation result

learn.py measures everything but credit, brush and the movement, camera and backgrounds
knobs, which it keeps from an existing style.json; fill the knobs from viewing the source and mark
them "basis": "judgment". check is learn.py's CHECK, the statistics shown to separate, less any
content-class row the source defines in fewer than two parts; measured_only lists the
subject-sensitive statistics the check prints but never judges. The
numbers live in style.json only; STYLE.md and other docs point at its keys.

Loop

1. Controls: controls.py measure <out> --near <a near-miss film> --other <films in other styles> --source <the clip> --replica <rotoscope work dir> measures a filtered family of the near miss
   (blur, noise, stripes, dry brush, contour warp), synthetic flat polygons, the other styles, and
   the same-style positives. It alternates the controls between a calibration and an evaluation
   half.
2. Learn: learn.py <work> <pack dir> --cuts <the clip's shot starts> --credit "<who made the source, its URL, date>" --negative <out>/calibration/*.json. Bands come from film-sized
   excerpts of the source, split by shot, with half the splits held out. The calibration controls
   cap each band's widening so they keep failing.
3. Prove the check: controls.py check <out> <pack dir> must show every control failing (margin
   above 0) and the source and replica passing, with the evaluation half held out. Then
   controls.py selftest <pack dir> --near <the near-miss film> (the pack argument is required)
   must exit 0. Put the table in reference/statistics.md. A control that passes goes under "what the check cannot
   say"; never narrow a band until it fails.
4. Describe: fill the judgment knobs in style.json and write STYLE.md.
5. Validate: author a scene of at least a third of the source's length (bands are learned from
   excerpts that long) on a different subject, from ink.js and the scene elements, with geometry you
   write, never traced; pass the pack's palette.ink as every ink.js call's color (the engine's
   own INK default is not the pack's). Render it at the source's size with ${CLAUDE_PLUGIN_ROOT}/scripts/render.py <scene.js> <frames dir> --fps <the pack's knobs.frame_rate.base_fps> --encode mp4 --playwright-core '${user_config.playwright_core}' (it serves the scene's folder and scripts/), and run inkstats.py <frames dir>.mp4 --cuts <shot starts> --pack <pack dir>. A produce production passes its shots.json as --cuts instead of retyping
   the starts. It exits 0 only when every checked row and the palette pass. Each row prints
   its distance to source: 0 is the source value, 1 the band edge on either side. The film's
   distance ranks passing scenes and its margin names the row nearest failing.
6. Review: measure each prop with --region X,Y,W,H --t T0-T1 next to source prop boxes, then
   read every drawing at 1:1 (tile frames so no image is over the 2576 px long edge
   ${CLAUDE_PLUGIN_ROOT}/skills/rotoscope/reference/method.md, Reviewing images, records; never
   downscale, never a GIF) beside source crops. A pass on the numbers with a crop that does not
   read as the style is a fail; say which mark is wrong.
7. Record: put the parameters that passed into the pack's brush and the result into STYLE.md,
   including where the scene still falls short.

Repeat 5-7 until the check passes and the crops read as the style. Tune the scene against the
check, never the check against the scene. inkstats.py also measures any film with no pack
(--json for the full summary), which is how to compare two films.

Next

/animation:produce

It takes a style pack this skill wrote and makes the film: boards for approval, then shots, render, and review.

Gotchas

• A post filter must not be what passes the check. Blur, noise, stripes and dry-brush overlays move
  edge softness and the amount of gray inside the ink, so those rows are measured but not checked.
  grain and period are measured only: on the woodcut source they follow codec residue and the
  pixel grid, and a flat black through h264 lands in grain's band. flat rejects texture laid
  over all the black, and rough and offstep do not move under blur, noise, stripes or dry brush
  at all. controls.py selftest <pack dir> [--near <film>] exits 1 if the review's blur-and-stripes
  filter (and, with --near, the warped, retimed ATTACK) gets a synthetic or near-miss film
  through; rerun it with --near after changing a statistic.
• straight and sliver follow the subject, so they are checked only inside the content classes
  border and caption (straight_border, sliver_caption, ...) and printed as measured only over
  the whole frame. boil is measured on the border and caption. A scene without a frame line or a
  caption panel in the top quarter leaves those rows n/a; draw both, as the source does.
• Bands are learned per excerpt of the whole source, not per shot, so a very short film is judged
  against bands that may not fit it; there is no per-shot row.
• w50 and pw50 come from a distance transform and move in 2 px steps; every woodcut source
  excerpt has pw50 8, which is why neither is checked.
• A strictly regular timing fails offstep: hold a few drawings for 2 or 4 frames.
• soft needs a soften pass after drawing, since a two-tone canvas render has almost no edge ramp.
  Chromium's ctx.filter = 'blur()' does nothing below a floor
  (${CLAUDE_PLUGIN_ROOT}/skills/rotoscope/reference/method.md, Rendering, holds the dated record); soften in JS instead.