design-intelligence

design-intelligence

The grounded source for frontend design decisions (ADR-061; first consumer of corpus-grounding). When the UI directive set's design step emits @agent-directive: ui-design-brief, consult this corpus FIRST and pre-fill the brief candidates — then the human confirms (design_confirmed). Corpus output is a constraint set, never final microcopy; the placeholder lock in directives/ui/design.ts is unaffected.

Boundary (council-locked): design.ts stays a pure orchestration gate and never imports the engine — the corpus call lives HERE, in the skill layer, keeping the engine an optional dependency.

Corpus: 11 tabular CSVs under data/ (161-row ui-reasoning decision map, WCAG-adjusted color token sets, 84 styles, 73 font pairings, 25 chart rules, UX/react/mobile guidelines) + 16 prose design-language specs (references/design-languages.md)

• the 10-category pre-delivery checklist (references/design-rules-checklist.md). Provenance + licenses: ATTRIBUTION.md; manifest: data/manifest.json.

When to use

• The UI directive set emits @agent-directive: ui-design-brief (the design step found state.ui_design empty).
• Any pre-build selection question: which style / palette / font pairing / layout pattern / chart type / icon system fits this product.
• Stack-idiom lookup before writing UI code (--stack axis).

Cross-task design memory — read DESIGN.md / PRODUCT.md first

Before running the corpus grounding or producing any design brief, check the project root for DESIGN.md and/or PRODUCT.md (written by design-system-capture). If they exist:

1. Read DESIGN.md — apply its captured visual decisions (radius, shadows, motion, spacing) as project constraints that take precedence over corpus suggestions. The corpus fills gaps; DESIGN.md overrides.
2. Read PRODUCT.md — note interaction patterns that affect the design (e.g., destructive-action policy, empty-state approach) so the brief is consistent with existing product conventions.
3. After generating the design brief: if a decision was made that isn't yet in DESIGN.md (e.g., chose a specific shadow for a new elevated surface), flag it for capture: "Suggest adding to DESIGN.md: elevated surface shadow = …"

Boundary vs brand-to-tokens/.tokens.json:

• .tokens.json = primitive definitions (gray-700 = #374151)
• DESIGN.md = usage decisions (elevated surfaces use the gray-700 shadow, 8px radius) Both are consumed; DESIGN.md takes precedence for usage questions.

Register — brand vs product

Determine the design register before grounding (see docs/guidelines/design-modes.md): brand mode ("the impression IS the product" — marketing, landing, consumer first-impression) prioritizes distinctive selection; product mode ("design serves the task" — dashboard, admin, workflow) prioritizes earned familiarity and accessibility. The register changes which corpus selections are appropriate (distinctive palette/typography in brand mode; predictable, semantic in product mode). State the register in the Design Read line below.

Embedded vs standalone (a third discriminator). UI embedded inside a host surface — a widget in a slide, a card in a chat, a panel in someone else's app — follows a flatter charter than a greenfield standalone page: restrained weights, hairline borders, no atmospherics/gradients/shadows that would fight the host. Select the register per surface (embedded → flat; standalone → the brand/product register above) — it is a selector, not a fixed token set (no values vendored; the host's tokens win).

Design Read — articulate intent before generating

Before producing any design brief or making any style selection, emit one line that declares the design read:

Reading this as: <page-kind> for <audience>, <vibe> language, leaning <design-system>.

Examples:

• Reading this as: SaaS dashboard for internal ops teams, functional language, leaning Radix/shadcn.
• Reading this as: marketing landing for B2C consumer product, playful editorial, leaning custom tokens.
• Reading this as: admin panel for technical users, dense/utilitarian language, leaning data-grid primitives.

If context is incomplete: state so and proceed exploratory — "Design context incomplete: no audience defined; proposing exploratory direction, expect revision after audience is clarified." Do NOT block on missing context; do NOT prompt the user with a gate; state the gap and continue.

Taste Dials — quantify, infer, emit

If DESIGN.md declares ## Taste Dials, use those values. Otherwise infer three 1–10 dials from the brief and append them to the Design Read line (… · dials V/M/D = 6/3/4) so the user can correct them; on confirmation, suggest persisting to DESIGN.md (via design-system-capture). Dials are a config, not a vibe — never re-infer when DESIGN.md already sets them (no drift across sessions).

Dial Inference Table (brief signal → Variance / Motion / Density, 1–10):

Dial → downstream levers (how a dial value changes generation):

Dials persist in DESIGN.md; the stack executors (tailwind-engineer, react-shadcn-ui, blade-ui, flux) read DESIGN.md and honour them.

Anti-Default Discipline — first-impulse check: Before committing to any design direction, cross-check your first impulse against the current-generation tells in design-antipatterns.md (§ Current-generation tells — the warm-editorial C5+T2+T7 signature and the previous-generation C1/C2 gradient) plus the L1/L2 layout defaults. If a tell was your first reach, name a different direction or explicitly justify why this brief genuinely calls for it. (The finalization cross-check against the full catalog is under Anti-slop discipline below.)

Honesty / real-system grounding

When the brief maps to an official design system (Material Design, Fluent, Carbon, Polaris, GOV.UK, shadcn, Tailwind UI, Radix, etc.):

0. Canon grounding first. If the brief names a system OR components.json/deps signal one (@mui/material, antd, @fluentui/*, @carbon/*, @atlaskit/*), pull docs/guidelines/design-canon.md, surface the matching one-line summary, and offer to fetch the live spec before committing to the system's conventions — rather than improvising. The canon index is thin + lazy: do not load it for a generic, unnamed brief.
1. Install the real package — do not hand-recreate its CSS or components. Surface the install command for the project's package manager (the system's official package, e.g. the shadcn CLI or the @mui/material distribution) and link the canonical documentation URL.
2. Never label an approximation as the official system. If generating approximate CSS for a system the project does not yet depend on, label it explicitly: "Approximation of Material Design elevation — not the official @mui/material package; install the package for production use."
3. If no official system is relevant: pick a deliberate creative direction (see Design Read above); never fall back to an unnamed generic aesthetic (per source-discovery-gate: real source before guessing).

Grounding precedence (consistent with brand-source-of-truth): consumer brand tokens > confirmed session decisions > named canon (design-canon.md) > generated corpus. Canon is a gap-filler, never an override of a registered brand value.

Procedure: Produce a grounded design brief (ui-design-brief rebound)

1. Ground (one call — engine runs the manifest's plan product → style → color → landing → typography with decision rules):

   ./scripts-run <skills-root>/corpus-grounding/scripts/ground ground \
     --manifest <skills-root>/design-intelligence/data/manifest.json \
     "<product type + mood + platform>" --json
   

2. Translate selections into the brief for state.ui_design:

   • layout ← landing/pattern selection (Section Order, CTA placement)
     • the reasoning rule's Recommended_Pattern;
   • components ← audit reuse first (existing-ui-audit inventory wins over corpus suggestions — never propose a new component the audit already has);
   • states ← required five (empty/loading/error/success/disabled), styled per the selected design language;
   • microcopy ← agent-written, final strings — the corpus never supplies microcopy;
   • a11y ← color selection's contrast-adjusted token set + the checklist's CRITICAL rows + accessibility-auditor method;
   • style/typography/effects/anti-patterns ← the grounded selections verbatim, with alternatives listed.

3. Always surface the grounded output's confidence + evidence_gap lines in the brief summary — the user signs off on what the corpus could NOT support, not only on what it could.

4. On design_confirmed: true the directive engine advances; revisions loop back here.

Font fallback (no google-fonts index — by design)

The 745 KB Google-Fonts index was rejected (ADR-061 §8): it duplicates a public API. When a requested font is outside font-pairings-reference.csv's 73 pairings: query https://fonts.google.com/specimen/<Family> (or the webfonts API) for metadata, OR propose the nearest curated pairing and say why. Never invent pairing metadata.

MASTER.md + page overrides ↔ state.ui_design (mapping)

The upstream cross-session memory pattern maps onto our delivery state:

File persistence stays opt-in (ground … --persist <dir>) and writes under the consumer's project as a durable artifact for multi-session consistency; on a fresh session, re-hydrate state.ui_design from MASTER.md + the page file before re-running design.

Grounding the review/polish a11y gate (charts + contrast)

The review/polish steps gate on state.ui_review.a11y. Ground two finding classes instead of ad-hoc judgment:

• Chart-type findings — the grounding CLI (ground via ./scripts-run) — …/ground search --manifest … --domain chart "<data shape>" → Accessibility Grade, A11y Fallback, Color Guidance columns justify "wrong chart type / missing colorblind fallback" findings with a citable row.
• Contrast findings — --domain color "<product>" returns the WCAG-adjusted token set; a finding that a hex pair deviates from the adopted set cites the row instead of eyeballing ratios. Auditing method stays with accessibility-auditor.

Stack guidance (--stack axis)

Per-framework Do/Don't corpora (16 stacks) ride the same manifest:

./scripts-run <skills-root>/corpus-grounding/scripts/ground search \
  --manifest <skills-root>/design-intelligence/data/manifest.json \
  --stack react "list rerender memo" [--filter "Severity=HIGH"]

Stack executors (blade-ui, livewire, flux, react-shadcn-ui, tailwind-engineer) pull idiomatic guidance + docs URLs from here instead of memory.

Output format

1. Grounded brief candidates per state.ui_design slot (layout, components, states, microcopy placeholder-free, a11y) — selections cited per corpus row.
2. The grounded output's confidence label + every evidence_gap line, verbatim, in the brief summary.
3. Alternatives list per domain so the human can swap before design_confirmed.

Do NOT

• Do NOT let the corpus write microcopy — it supplies constraint sets; final strings are agent-written (placeholder lock stays in force).
• Do NOT import the engine into directives/ui/design.ts — council boundary; the corpus call lives in this skill layer.
• Do NOT propose a new component the existing-ui-audit inventory already covers — audit findings outrank corpus suggestions.
• Do NOT hide low confidence — the user signs off on the gaps too.

Diagram-type routing — route on the verb

Choose a visualization by the intent verb, not the noun. Count the nouns before you draw (input-complexity triage): 1–2 → inline prose or a single shape; 3–7 → one diagram; 8+ → split or summarize, never one dense picture.

Geometric pre-checks (run BEFORE finalizing an SVG/diagram)

Ranked by failure rate — procedures, not constants:

1. viewBox safety — compute the lowest + rightmost element (plus a buffer) and set the viewBox from that; never assume the default fits.
2. arrow-through-box trace — trace every arrow's path and confirm it does not cross through an unrelated box before drawing it.
3. box-width-from-longest-label — size each box from its longest label before placing it, so text never overflows.

(Reference-only: any color/easing/frame values come from the consumer's tokens or a maintained upstream — this skill vendors no drawn-asset corpus.)

Interplay (who owns what)

Gotchas

• Corpus grounds pre-action selection — do not use it as mid-task reference (open references/ instead) or as a validator (rules own that).
• Empty result ≠ error: surface the evidence gap and proceed on priors.
• Keep queries product-shaped ("fintech dashboard", "luxury e-commerce mobile") — the detect map routes generic words to style.

Anti-slop discipline

Before finalizing any design brief, cross-check against docs/guidelines/design-antipatterns.md — especially the Color (C1–C5), Typography (T7–T8), and Layout (L1–L2) sections. If the grounded corpus selection lands on a pattern in the catalog, either invoke the override condition or adjust the selection. Run the AI-slop originality self-test (catalog § "The AI-slop originality self-test") on the chosen aesthetic direction before emitting design_confirmed.

Why this skill is rich

This skill carries 11 tabular CSVs (161-row UI-reasoning decision map, WCAG-adjusted color token sets, 84 styles, 73 font pairings, 25 chart rules, UX/ react/mobile guidelines) plus 16 prose design-language specs and a 10-category pre-delivery checklist. Agents need to see the full corpus to make grounded selections — condensing to a summary destroys the evidence trail ("corpus row 47 justifies the palette choice") that the skill's output contract requires. Compressing the 16 design-language specs into fragments makes the style selection unreproducible and audit-unfriendly.

Policies

• Upstream MIT + Apache-2.0 obligations: ATTRIBUTION.md.
• Refresh: quarterly per the manifest; bump upstream.last_checked on every refresh (ADR-061 §6).