Structure engine (phase 1): named macrostructures + archetype catalog + diversification

[codeswithroh]

Background — studied Hallmark

Studied Nutlope/hallmark (15.7k★, by Together AI), the strongest anti-slop design skill in the field, end to end — its 558-line router SKILL.md, ~90-file reference tree, 50-archetype component cookbook, 58-gate slop test, and token architecture.

Its core mechanism is a combinatorial catalog of named page shapes + component archetypes with stable IDs, plus a project-memory diversification engine that forces each build to differ structurally from the last. That's the thing we didn't have — and it closes our single biggest gap.

The gap this closes

tastemaker already varies color per project (the palette generator + contrast contract). But our structure was monocultured: every landing page reached for the same hero → 3-feature-cards → CTA → footer rhythm. That's the strongest "an AI built this" tell at the page level, and it survives a perfect palette — the deepest unaddressed part of the #9 monoculture concern. Color varying while structure repeats still reads as two color-swaps of one template.

What this does — and what it deliberately keeps

Grafts Hallmark's structural machinery onto our foundation while keeping the two axes where we already beat Hallmark: the per-project generated palette + contrast contract, and the real-asset pipeline (Hallmark is typography-only with placeholders). So the archetypes here are asset-forward — real mockups/charts/photos/illustrations, not typography-only parts.

New reference files:

• references/macrostructures.md — 12 named whole-page shapes (Feature Stack, Editorial Index, Long-Scroll Narrative, Bento Showcase, Stat-Led, Gallery Grid, Product Demo, Split Diptych, Conversational FAQ, Manifesto, Catalogue, Poster Fold) with mood affinity + an honesty rule binding proof-led shapes.
• references/component-catalog.md — named archetypes with stable IDs + variation knobs (Nav N1-N7, Hero H1-H6, Feature F1-F7, Proof P1-P4, CTA C1-C4, Footer Ft1-Ft6, Section-head S1-S4), per-mood routing, cross-cutting rules (no fake chrome, one icon family, banned eyebrow-left head).
• references/diversification.md — the project-memory engine: .tastemaker/log.json schema, the rotation rule (macrostructure/nav/footer/hero each differ from last build; change a knob on reuse), the "state it out loud" accountability step, the CSS build stamp. Palette is explicitly exempt (already varies by generation) — so our diversification is purely structural, a simplification our generator earns vs Hallmark's multi-axis theme rotation.

Wiring: SKILL.md gains Step 2.5 (pick structure + diversify); Step 0 reads the log, Step 4 writes the stamp + log entry; reference table updated. App-shell screens explicitly skip this (their shape is the sidebar+topbar frame). style-lock-format.md gains a Structure section; anti-slop-checklist.md gains two structural-variety gates; component-patterns.md's Landing section is cross-linked as the finer-grained guidance inside a macrostructure.

Phased plan (this is phase 1)

• Phase 1 (this PR): structural-variety engine.
• Phase 2 (next): upgrade the prose anti-slop checklist into a numbered, mood-scoped slop-gate list + a pre-emit self-critique pass.
• Phase 3: study (extract design DNA from a URL/screenshot) and audit verbs.

Test plan

• All cross-references between new + existing files resolve (grep check)
• Fence balance and heading structure verified in all three new files
• SKILL.md step order reads 0 → 1 → 2 → 2.5 → 3 → 4 → 5
• No stray "preset" mentions introduced
• Kept our generator + asset pipeline intact — this is additive structural machinery, no regression to color or assets

🤖 Generated with Claude Code