Personalized Kids' Coloring Book

Produce a letter-size vector PDF of bold line-art pages starring the user's kids/family/pets.
Characters are PARAMETRIC (trait vectors), never photo-likeness generation — this guarantees
the same recognizable character on every page and avoids identity drift entirely.

NON-NEGOTIABLE: this skill's quality lives in executing lib/charlib.py with Python
(cairosvg + ghostscript). Never re-implement the drawing library by hand or "approximate"
its helpers in another language/canvas — hand-reimplementations lose the white-fill
back-to-front layering that makes overlapping shapes read as solid objects (you get
transparent outline soup: trees as crossing circles, paths through heads). If you cannot
execute Python in your environment, STOP and tell the user that instead of improvising.

FIRST: read reference/model-tiers.md and pick your operating mode honestly
(full-creative / standard / conservative) based on your own capability tier. It also
contains the universal pre-flight layout plan — every model bottom-crams first drafts
without it — and the escalation ladder for when an element won't come right after two
attempts (compromise → delegate the element to a stronger model → recommend escalating
the whole build).

Workflow

1. Cast setup — trait vector per character

For each kid collect/derive: name, age, hair style (one of charlib.HAIR_STYLES:
bob_bangs, bob [+headband], tousled, long_wavy, pigtails, curly, buzz), glasses y/n,
freckles y/n, a personal motif (star/strawberry/truck/butterfly/dino/ball...), favorite
things, relative height. Pets: species (dog/cat), coat (plain/spots/patch/stripes), collar.
Record the OUTFIT in the trait dict and use it everywhere (t["outfit"]) — never
override per-page, and when a cast recurs across books, copy trait dicts verbatim
(silently changing a signature outfit or feature breaks recognition).
Adults (grandparent etc.): long_wavy hair + glasses works well.

If PHOTOS are provided: look at them ONLY to classify trait buckets (hair style/length,
glasses, freckles, pet coat pattern) — never attempt to reproduce a face. State the traits
you extracted so the user can correct them. Distinguishability at this fidelity comes from
hair silhouette + glasses + freckles + motif + height, not facial geometry.

2. Mode + plan

• themed (default): ~10 pages, each starring the kids together in one child's favorite
  activity; rotate whose interest leads each page; every kid appears on most pages.
• story (ages 3-6): use the 10-beat arc in reference/story-mode.md. If the user
  hasn't specified a story, offer 2-3 tropes from reference/story-tropes.md matched to
  the child's age and interests (one-sentence pitch each) and let them pick. Captions on
  every page: 15-25 words, simple present tense, repeat names, chant-along refrain every
  1-2 pages, the child acts (never watches). Low-stakes obstacles only (hidden toy, shy
  pet). No rhyme.
  Plan page-type variety: cover, name-tracing page (large outline letters of the kids' names),
  scene pages, one find-the-X activity page, back cover ("Made with love for ...").

3. Render

Write a make_book.py in the output dir that copies the proven structure:

import sys; sys.path.insert(0, "<skill_dir>/lib")
from charlib import *
def cover(): ...
BUILDERS = [("01-cover", cover), ...]
if __name__ == "__main__":
    import sys as s
    build(BUILDERS, OUTDIR, only=s.argv[1:] or None, pdf_name="<Kids>-Coloring-Book.pdf")

Use charlib primitives + figures (kid_stand, kid_sitting, dog, cat_sitting,
face_traits) and motifs; draw scene furniture with P/C/E/LINE. Read
reference/drawing-guide.md FIRST — it encodes the collision gotchas and age-band line
weights. Keep pages deterministic; partial rebuild via python make_book.py 03 must work.

For everyday objects that have no helper (car, truck, tractor, train, boat, tree,
barn, animals, playground gear...), do NOT invent complex geometry freehand — LLMs
reliably botch object proportions. First check charlib for a helper
(60+ shapes: vehicles, buildings incl. castle_small/treehouse/village_house,
animals incl. dog_sit/dog_sleep/reindeer/lion/baby_goat, furniture, food,
costume props like chef_hat/bib_apron/tiara, scene props like tent/campfire/
ferris_wheel/lamppost/wardrobe — grep charlib for ^def); if none fits, follow the
relative-proportion recipes in reference/shape-cookbook.md and build the object
from primitives at your chosen scale. Wrap each placed object in matted() over any
background. See CREDITS.md for the recipes' provenance.

4. Self-inspect + QA loop

After building, Read every page PNG yourself and fix obvious collisions/floaters. Then run
tile QA: render_tiles(svg, qa_dir, name) per page and review the overlapping tiles (spawn
sonnet subagent reviewers if the Agent tool is available, one per page, prompt them for:
unintended overlapping lines, tangent-line/figure-ground illusions where background
contours touch figures, floating/disconnected parts, out-of-proportion elements,
ambiguous shapes, eerie/off-putting facial expressions, elements cut off;
severity-tagged findings). Apply fixes, re-render, and
re-verify the pages that had HIGH findings. 1-2 fix rounds is normal.

5. Deliver

Merge to PDF via build() (cairosvg + ghostscript). Report the PDF path, page list, and
which traits distinguish each character. If the user has a family share, offer to copy there.