visual-grammar-selector

name	visual-grammar-selector
description	Select varied tasteful visual grammars for HTML-first academic and technical decks so outputs do not look like one repeated template.

Visual Grammar Selector Skill

Pick a visual grammar before rendering. Do not default every deck to the same palette, layout cadence, or typography rhythm.

Available Grammars

• editorial-profile: public researcher, lab, or portfolio profile.
• swiss-systems: systems, architecture, benchmarks, infrastructure.
• dark-lab-notebook: frontier AI, experiments, agent traces, dense research notes.
• paper-atlas: paper reading, survey, citation-heavy academic decks.
• keynote-evidence-wall: demos, launch-style research talks, visual proof-heavy pitches.
• highsense-gallery: high-end web editorial feel for profile or product-research decks.
• academic-homepage-grid: PRISM-like academic homepage structure for dense researcher/project profiles.
• prism-dossier: polished academic homepage dossier with identity rail and selected-work flow.
• fathom-research-brief: scientific information-design memo for benchmarks, systems, and evaluation.
• jetset-theory-grid: strong typographic grid for conceptual theory and method framing.
• monograph-review: PRISM + Information Architects academic monograph style; dense, source-linked, serif/sans, no decoration.
• broadside-lab: Huashu Broadside / editorial manifesto style; dark, typographic, single fire accent, useful for strong research claims.
• catalog-atelier: curated research-catalog style; warm paper, sage/terracotta, gallery rhythm, good for people/project evidence decks.
• evidence-atelier: curated proof-artifact wall; warm paper, terracotta, irregular bento/index pages.
• atlas-marginalia: annotated academic atlas; near-white page, blue source accent, plate plus marginal note rail.
• systems-field-manual: operational systems/benchmark manual; hard grid, ledger rows, one orange signal color.
• lab-trace-ledger: dark experimental trace ledger for agent/RL/reasoning behavior decks.
• object-study-wall: gallery-like object/source study wall for project and public artifact profiles.
• forest-editorial-brief: warm editorial forest profile grammar; polished academic without template flavor.
• cobalt-research-grid: hard cobalt technical grid for systems, benchmarks, and tool-use evidence.
• vellum-research-note: formal dark research-note grammar for seminars, thesis, and paper defenses.
• mono-ink-ledger: monochrome dossier/handout grammar for source-linked reviews.
• neo-grid-lab: high-energy neon grid for demos, AI products, and engineering showcases.
• prism-clean-room: PRISM-like clean academic homepage grammar with denser publication/profile flow and no card gloss.
• prism-publication-stack: publication-forward academic homepage stack for people/lab/project profiles.
• prism-newsroom-index: PRISM-like academic news/index grammar for selected work, papers, and public-source profiles.
• ia-research-archive: Information Architects-style source archive with marginal rails, blue source links, and dense text-first proof.
• broadsheet-data-room: dense broadsheet/data-room grammar for benchmark reviews, paper dossiers, and source-heavy profiles.
• pentagram-gridnote: bold typographic gridnote for methods, claims, and systems talks that need poster-level confidence.
• takram-research-system: calm research systems memo with precise grids, thin rules, and quiet proof-first evidence.
• stamen-data-map: cartographic/data-map grammar for project, benchmark, geography, or public-source evidence decks.
• couture-exhibition: high-end editorial exhibition grammar with large proof plates and wall-label captions.
• js-editorial-cascade: high-sense web editorial cascade with asymmetric proof staging and strong type/image contrast.
• sumi-research-scroll: ink-and-paper research scroll for source-indexed academic profiles and marginal proof reading.
• huashu-editorial-lab: Huashu-inspired typographic lab grammar with hard grid, one accent, and anti-cliche discipline.
• huashu-build-board: Huashu Build-style hard benchmark board with oversized metrics and black/red rule structure.
• huashu-issue-broadsheet: Huashu-style issue broadsheet for bold technical profiles, benchmarks, and handout-like evidence boards.
• gallery-proof-room: warm gallery proof-plate grammar for artifact-rich public profiles and project evidence decks.
• prism-workbench-index: PRISM-inspired researcher workbench with selected-work rows and compact source-index pages.
• signal-intelligence-brief: institutional briefing grammar with source ledgers, serif authority, and restrained signal accents.
• raw-grid-research: neobrutalist research board for benchmarks, demos, and methods that need hard-framed confidence.
• stencil-field-tablet: archival field-manual grammar with colored tablets for mapped projects, surveys, and source specimens.
• maison-research-catalog: high-end catalog grammar for public profile/project decks with dominant source plates.
• folio-swiss-noir: black/white Swiss folio grammar for severe method, benchmark, and source-exhibit decks.
• chromatic-research-map: chromatic map grammar for benchmark terrain, research geographies, and multi-threaded public evidence.

Selection

• High-taste public researcher/profile defaults: forest-editorial-brief, object-study-wall, or evidence-atelier.
• Conservative public researcher/profile default: academic-homepage-grid or prism-dossier.
• People/profile fallback: editorial-profile.
• Tooling, systems, benchmarks: systems-field-manual; use swiss-systems as a quieter fallback.
• RL, agents, frontier model work: dark-lab-notebook.
• Literature review or paper talk: paper-atlas.
• Demo or product-research story: keynote-evidence-wall.
• When the user asks for "高级感" or a less templated public profile: highsense-gallery.
• When the user asks for academic personal homepage taste or fuller research-profile layouts: academic-homepage-grid.
• When profile decks feel too sparse or résumé-like: compare academic-homepage-grid and prism-dossier.
• When PRISM-inspired decks still feel too much like cards or résumé blocks: compare prism-clean-room.
• When the profile has enough papers/projects to behave like a homepage publication list: compare prism-publication-stack.
• When a profile needs news/index density rather than a résumé arc: compare prism-newsroom-index.
• When a deck should read like a serious source archive rather than a talk template: try ia-research-archive.
• When evidence and comparisons should read like a newspaper/data room: try broadsheet-data-room.
• When claims depend on measurements, benchmarks, or source-linked evaluation: use fathom-research-brief.
• When the deck needs a bold reset and can survive a sharper typographic voice: use jetset-theory-grid.
• When the deck needs a poster-grade typographic reset without neon or gradients: try pentagram-gridnote.
• When a technical story needs calm product/research-system taste: try takram-research-system.
• When the content is project geography, benchmark terrain, or public web evidence: try stamen-data-map.
• When the deck needs high-end web/editorial atmosphere without AI gloss: try couture-exhibition.
• When the user explicitly wants HTML taste translated into slides, compare js-editorial-cascade for web-editorial asymmetry and sumi-research-scroll for a quieter ink-scroll academic reading.
• When the deck needs Huashu-style typographic confidence and hard critique discipline: try huashu-editorial-lab.
• When benchmark/system claims need a hard Build-style board: try huashu-build-board.
• When a profile or benchmark deck needs a bolder issue-board treatment without gradients: try huashu-issue-broadsheet.
• When source artifacts are the strongest material and should dominate: try gallery-proof-room.
• When PRISM-style decks feel card-like and need a selected-works index instead: try prism-workbench-index.
• When a profile feels too card-like or too ordinary: compare monograph-review and catalog-atelier.
• When a systems/profile deck needs stronger stage presence without gradients: try broadside-lab.
• When a deck has strong source artifacts but feels pasted together: compare evidence-atelier and object-study-wall.
• When a profile needs warmth and editorial maturity without becoming decorative: try forest-editorial-brief.
• When systems evidence needs more technical bite than systems-field-manual: try cobalt-research-grid.
• When the deck should feel like a formal research note or defense: try vellum-research-note.
• When the output is closer to a dossier or handout than a stage talk: try mono-ink-ledger.
• When the brief is demo/product/engineering and can tolerate a loud grid: try neo-grid-lab.
• When paper/profile material needs source annotation over stage presence: try atlas-marginalia.
• When tool-use or benchmark material looks like a generic summary: try systems-field-manual.
• When agent/RL material needs a dark research mood without generic neon AI flavor: try lab-trace-ledger.

Current Judging Bias

The historic strict bake-offs across Fixture A and Fixture B fixtures cleared zero blocking layout errors, but renderer-wide safety rules have since become stricter. The default pool is now 42 grammars after adding signal-intelligence-brief, raw-grid-research, stencil-field-tablet, and the Maison/Folio/Chromatic high-sense triad. The latest focused layout-contract pass cleared gallery-proof-room, js-editorial-cascade, raw-grid-research, stencil-field-tablet, and huashu-issue-broadsheet across the two fixtures with populated Image Contracts, zero blocking errors, and zero strict warnings. The latest overlap/image repair also cleared academic-homepage-grid on both Fixture A and Fixture B after compact proof layouts stopped rendering callout-note lists and source-heavy artifacts were enlarged to roughly 1000x556 visible source pixels. It ranks forest-editorial-brief, object-study-wall, evidence-atelier, and cobalt-research-grid as the highest-taste default pool, while academic-homepage-grid, prism-dossier, prism-clean-room, prism-publication-stack, prism-newsroom-index, and prism-workbench-index remain conservative or publication-forward profile defaults. Treat vellum-research-note, broadsheet-data-room, mono-ink-ledger, atlas-marginalia, monograph-review, paper-atlas, catalog-atelier, ia-research-archive, and fathom-research-brief as formal/dossier special cases. Treat neo-grid-lab, jetset-theory-grid, keynote-evidence-wall, and broadside-lab as deliberate resets, not defaults; pentagram-gridnote and huashu-editorial-lab, huashu-build-board, and huashu-issue-broadsheet are more formal typographic resets, while couture-exhibition, gallery-proof-room, and js-editorial-cascade are editorial proof-plate or web-editorial resets, and sumi-research-scroll is an ink-scroll source-index reset. Avoid selecting a grammar only because its palette differs from another variant.

Warnings affect ranking even when they do not block export: proof-image-small and proof-image-rendered-small mean the source pixels are underfeatured, artifact-image-rendered-small means a source panel will not read at projection scale, callout-overlap means the annotation layer needs repair, proof-caption-missing / proof-caption-generic / artifact-caption-generic mean the image lacks source context, caption-text-too-small / caption-contrast-low mean the source note is not readable enough, label-contrast-low means support labels fit but are not visually legible, and decorative-image-too-small means a cover image is only an identity cue, not evidence. title-wrap-deep and subtitle-wrap-deep mean the copy is close to using the slide like a document; shorten or split before delivery. In strict export, these warning gates fail --fail-on-layout.

Read layout-audit-report.md before trusting any shortlist. The Image Contracts table should show proof and artifact images with enough rendered source pixels, useful visible area, healthy slot use, and crop/caption/source status. A visually polished grammar is not ready if the table shows a source surface as a small identity cue or an uncropped decorative image. For repo, UI, paper, table, benchmark, project-page, or homepage screenshots, look for about 980x540 rendered source pixels or better. If the source-heavy surface falls below roughly 820x460, repair the crop or slot before judging style.

Before finalizing a selection, reject empty modules and repeated artifacts: labels, metrics, bento cells, and evidence boxes need real content; the same homepage/screenshot should not appear more than twice unless the crop and slide job change.

Renderer truncation is a source error, not a layout trick. Do not submit more bullets, metrics, labels, or callouts than the renderer can display; split the slide, convert to a matrix, or move detail into notes. A wide cover screenshot must use an evidence: crop/caption/source contract, not bare image:.

Use the diversified shortlist, not just the highest raw score. compare-grammars reports Shortlist Diversity with a style family and composition family for each ready grammar, and also writes DESIGN_DIRECTIONS.md with a 20+ school direction library. Read that direction report before picking a winner: choose a safe academic base, a proof-forward editorial route, and a bold reset for the showcase. A shorter shortlist is acceptable when the remaining clean variants are near-siblings; do not add another PRISM/homepage variant merely to fill a quota. Prefer one academic-homepage, one proof-gallery, one systems-grid, one formal-dossier, and one Huashu/bold reset direction when they are clean.

When the brief says the current output is too ordinary, too fixed in palette, or too template-like, start with:

uv run academic-deck compare-grammars --deck <deck.yaml> --out <out> --grammars highsense-20 --fail-on-layout

The highsense-20 preset is a curated reference pass, not a random sample: it mixes PRISM academic homepage structures, source archives, Beautiful HTML Templates families (signal, raw-grid, stencil-tablet), JS web-editorial proof staging, Huashu issue boards, gallery proof rooms, and the Maison/Folio/Chromatic triad. Inspect the stacked overview before choosing. In compare-grammars, --fail-on-layout should leave rejected candidates in the report and fail only when no clean shortlist variant remains. The final selected grammar still must pass single-route export with zero hard layout errors and zero strict image/density warnings.

For a tone-first template scan before grammar selection, use:

uv run academic-deck template-shortlist --brief "<occasion + mood + content>" --out <out>

After choosing, read the matching templates/<grammar>/DESIGN.md.

Composition Profiles

Do not judge variety from visual_grammar alone. Check layout-audit-report.md for named compositions such as:

• cover-source-rail: academic identity/source rail.
• cover-title-wall: large title with small source surface.
• cover-poster-grid: bold typographic cover.
• proof-stage: claim/sidebar plus dominant proof surface.
• proof-atlas-spread: proof-first spread with notes on the side.
• proof-dossier: homepage-style proof and source context.
• artifact-rail / artifact-dossier: normal content slide carrying a real source artifact.
• matrix-ledger: scan-friendly table/ledger composition.
• content-workbench-index: selected-work rows with a compact claim column.

A pass is not diverse if every candidate has the same composition sequence with only palette changes.