Atelier data

atelier-data — searchable design reference

A curated, local dataset + a tiny BM25 search engine. It answers "give me a vetted starting point / cross-check" fast, without spending model context on the whole corpus. It is the data layer of the Atelier suite; the other atelier-* skills remain the authorities that make the actual design decisions.

Provenance. Harvested and cleaned from UI/UX Pro Max v2.5.0 (© 2024 Next Level Builder, MIT). See references/PROVENANCE.md and ../../THIRD-PARTY-LICENSES/ui-ux-pro-max-MIT.txt. Content is curated-opinion (some generated), not a cited standard — treat it as informed starting material, not ground truth.

When to use (and when not)

Use it for: a fast per-industry palette/type/style starting point; product-type design reasoning (pattern + anti-patterns) during planning; a current do/don't lookup for Next.js / React / shadcn / Vue / Three.js while implementing or refactoring; a UX-heuristic rubric to hand a reviewer; a Google-Fonts metadata lookup.

Do not use it for: building the real token system (use atelier-foundations — its colors here are raw Tailwind hex stops, not OKLCH, and some fail AA); final font choice (use atelier-typography — this set still contains common defaults); auditing real code (use atelier-review — it cannot read your files); or as an always-on rule source (query it on demand only).

How to query (Windows: use python, not python3)

cd $CLAUDE_CONFIG_DIR/skills/atelier-data/scripts
python search.py "<query>" --domain <domain> [-n 3]
python search.py "<query>" --stack <stack>  [-n 3]
python search.py "<query>"                     # auto-detect domain

Keep -n small (≤5) and avoid --json for large domains — results enter the model's context, and --json bypasses the per-field 300-char truncation. Python is stdlib-only; no network calls.

Datasets → which Atelier skill owns the decision

Domain / file	Gives you	Decision owner
--domain product → products.csv (161)	Per-product-type style / landing / dashboard recommendations	atelier-direction, atelier-ux (planning)
--domain reasoning → ui-reasoning.csv (161)	Per-product-type recommended pattern, style priority, color/type mood, key effects, anti-patterns	atelier-direction, atelier-ux (planning)
--domain style → styles.csv (67 web; Mobile rows removed)	Style definitions, CSS-var stubs, AI-prompt keywords, framework scores	atelier-direction, atelier-components
--domain color → colors.csv (161)	shadcn-token palette per product type — scaffold only	atelier-foundations (rebuild in OKLCH + re-check contrast)
--domain typography → typography.csv (71; broken rows removed)	Font pairings w/ ready @import + Tailwind config	atelier-typography (authority on escaping defaults)
--domain google-fonts → google-fonts.csv (1,923)	Font metadata: category, axes, subsets, popularity	atelier-typography
--domain landing → landing.csv (34; uncited stats stripped)	Section-order patterns, CTA placement, color strategy	atelier-ux, atelier-layout
--domain chart → charts.csv (25)	Chart-type → library/a11y guidance	atelier-dataviz
--domain ux → ux-guidelines.csv (89 web/all)	Do/don't UX + a11y heuristics w/ code examples	atelier-perf-a11y, atelier-review (rubric feed)
--domain react → react-performance.csv (44)	RSC / React 18+ / Next 15 perf patterns	atelier-components
--stack nextjs|react|shadcn|vue|threejs	Current per-stack do/don't tables w/ doc URLs	atelier-components (threejs → atelier-webgl)

Honest caveats (read before trusting a result)

• Colors are Tailwind hex stops, partially a11y-checked. Use as a seed, then rebuild in OKLCH via atelier-foundations and verify contrast via atelier-perf-a11y. Never paste Secondary/Accent as body text without a contrast check.
• Typography still includes common pairings (Inter/Playfair etc.). atelier-typography is the authority and deliberately steers away from defaults — treat rows as options, not endorsements.
• ui-reasoning style picks can misfire (its Style_Priority can override better matches; the Decision_Rules JSON is advisory, not executed). Sanity-check against atelier-direction.
• landing.csv conversion %-claims were removed (they were uncited). The section-order/CTA advice remains.
• Counts/labels in the source's own marketing were inflated; the numbers above are the actual cleaned row counts.

What was deliberately left behind

The design-system generator (design_system.py), colors→hardcoded-CSS output, the MASTER.md/page-override generator, all mobile-native rows/stacks, html-tailwind (it was Tailwind v3), and the npm CLI — all conflicted with this stack or with Atelier. This skill is data + search only.