design— agent skill

Direction first, tokens second, code last. Restraint is the default posture; ONE intentional moment per surface earns the lift.

Posture

Direction precedes tokens; tokens precede code. The picked direction is the contract — palette, type, spacing, motion all derive from it, not the other way around. Restraint is the default; reach for decoration only when a named surface goal demands it. Posture rests on references/soul.md (design philosophy) — load when the user asks "why this look" or when the model is tempted to add decoration to compensate for a thin idea.

Balance, not maximalism. ODIN's "don't hold back" (borrowed from the <design> block above) means conviction, not volume. The two failure modes catalogued in the anti-slop charter below are reciprocals — AI-generic timidity (Side A) and decorative overkill (Side B) both come from the same fear of commitment. Balance simplicity and verbosity per surface; restraint as the default, the one intentional moment as the lift, neither timid nor loud.

ODIN <design> block — LOAD-BEARING

Quoted verbatim from output-styles/odin.md lines 254–262.

<design>
Modern, elegant UI/UX. Don't hold back.

**Tokens:** MUST use design system tokens, not hardcoded values.
**Density:** 2-3x denser. Spacing: 4/8/12/16/24/32/48/64px. Medium-high density default. Ask preference when ambiguous.
**Paradigms:** Post-minimalism [default] | Neo-brutalism | Glassmorphism | Material 3 | Fluent. Avoid naive minimalism.
**Forbidden:** Purple-blue/purple-pink | `transition: all` | `font-family: system-ui` | Pure purple/red/blue/green | Self-generated palettes | Gradients (unless explicitly requested, NEVER on buttons/titles)
**Gate:** Design excellence >= 95%
</design>

Register × paradigm matrix — pointer

Every design surface is brand (marketing, landing, campaign, long-form content, portfolio — design IS the product) or product (app UI, admin, dashboard, tool — design SERVES the product). Identify before designing.

Detection rule. First match wins:

1. Cue in the task itself — "landing page" / "campaign hero" → brand; "dashboard" / "settings panel" → product.
2. Surface in focus — the page, file, or route being worked on; the route segment usually disambiguates (/marketing/* vs. /app/*).
3. Register field in PRD or project context (AGENTS.md, CLAUDE.md, brand brief).

Brand register opens neo-brutalism / glassmorphism / M3 Expressive with Committed / Full palette / Drenched color strategies; product register defaults to Restrained + post-minimalism / Fluent 2. Full matrix in references/paradigms.md §0.

Anti-slop charter — LOAD-BEARING

Two-sided ban-list. Each row: ban — WHY in one line. Depth in references/anti-slop.md.

Side A — slop tells (the AI-generic look):

• Purple-blue or purple-pink gradient — RLHF over-aligns to this; betrays self-generated palette.
• Inter alone as the type system — default of every Vercel template; no commitment, no contrast.
• Centered hero plus 3-column feature grid — the shadcn landing-page silhouette; reads as preset.
• Glassmorphism on every surface — translucence loses meaning when nothing is opaque.
• rounded-lg uniform on every element — radius without hierarchy is decoration, not signal.
• shadow-md uniform across the surface — elevation that conveys nothing.
• transition: all — animates layout, color, and transform together; jank guaranteed.
• font-family: system-ui — abdicates the type decision; reads as "did not pick".
• Default Tailwind palette (slate-500 / blue-500) — the costume of "I used the framework defaults".
• Colored card borders to assert structure — borders are not the right tool for hierarchy.
• Emoji icons in production UI — accessibility hostile; locale-fragile; reads as draft.

Side B — overkill compensation (slop's louder cousin):

• Sprites on every empty pixel — decoration substituting for missing information density.
• Gradient on every section background — every section "important" means none are.
• Animation on every element entrance — motion budget is a budget; spend it once.
• Multi-paradigm mash (neo-brutalism shadow on a glass card on a Material 3 button) — paradigm conflict reads as confusion, not eclecticism.
• Decorative noise compensating for a thin idea — when the surface earns its weight, restraint amplifies it.

Color strategy axis (commitment level). Pick the strategy before picking colors. Each tier is deliberate, not a fallback.

• Restrained — tinted neutrals plus one accent at ≤10% surface coverage. Product default.
• Committed — one saturated color carries 30–60% of the surface. Brand default for identity-driven pages.
• Full palette — 3-4 named roles, each used deliberately. Brand campaigns; product data viz.
• Drenched — the surface IS the color. Brand heroes, campaign pages.

The "one accent ≤10%" cap applies only to Restrained. Committed / Full palette / Drenched exceed it on purpose; treat the cap as direction-conditional, not universal.

Direction-first workflow

Six steps. Do not skip the divergence step.

1. Frame the surface. Capture: surface (landing / dashboard / settings / docs / one-screen tool), primary user, density target, motion budget in ms. Then write one sentence of physical scene (who, where, ambient light, mood) that FORCES the dark/light decision. Category names alone do not force the answer — "observability dashboard" fails; "SRE glancing at incident severity on a 27-inch monitor at 2am in a dim room" succeeds. Pin why this design must feel a certain way before deciding how. Surface answers before generating directions; designing on assumed callers wastes the parallel budget.

1. Diverge: 3-4 directions in parallel via Verbalized Sampling. Dispatch one Explore agent per direction with a constraint that forces contrast (post-minimalism vs neo-brutalism vs Material 3 vs Fluent, or named taste anchors that pull in opposite directions). Reject converged outputs; re-dispatch with sharpened constraints if two directions read alike. See references/diversity.md for the six diversity-engineering techniques (verbalized sampling, actor-critic per candidate, persona injection, temperature, most-unlikely reframing, anti-pattern catalog).

1. Per direction, return a fixed shape. Each direction states: name (one or two words), 1-2 taste anchors (Linear / Stripe / Things 3 / Rosé Pine / Are.na — name the references), OKLCH palette stub (4-6 swatches, never the default Tailwind ramp), type pair (display + text, named families), spacing scale subset committed (e.g., 4/8/16/24/48), motion budget in ms with one easing curve.

1. Pick via per-axis single-select. Each axis (direction, density, motion budget, type pair) is its own single-select question; the recommended option carries (Recommended) and is placed first. Ticking (Recommended) is accepting the default. Never use multiSelect for axis-with-default override semantics — that shape collapses N independent decisions into one ambiguous checklist; reserve multiSelect for additive picks only (feature toggles, optional sub-tasks).

1. Derive DTCG-shaped tokens from the picked direction. Color, type, space, radius, shadow, motion — each a token, each referenced not hardcoded. Tokens precede component code; component code references tokens.

1. Route to the surface reference. Pick the row in §6 that matches the runtime; load that reference for surface-specific patterns and pitfalls.

Surface routing

Cross-runtime references

These references are runtime-agnostic — load them when the task fits the trigger, alongside whatever surface reference is already loaded.

Cross-surface invariants

• Color as input, never as default. Custom OKLCH palette derived from the picked direction; never the default Tailwind, Material, or Bootstrap ramp.
• Spacing scale is 4/8/12/16/24/32/48/64. Pick a subset that matches density target; commit and stick. A new value mid-build is a smell.
• At most two type families. Display plus text. A third family is a smell unless the direction explicitly demands it (e.g., a mono accent for code).
• Motion is budgeted in milliseconds. One easing curve per surface. transition: all is forbidden — name the properties (transition: opacity 120ms ease, transform 120ms ease) so layout and paint do not animate together.
• Semantic structure precedes class names. <nav> / <main> / <article> first; utility classes second. Class soup over weak structure is slop.
• Contrast: WCAG 2.1 AA is the legal floor. APCA is a design-input tool only — it was REMOVED from WCAG 3 in July 2023 and never reinstated. Do not claim "WCAG 3 compliant via APCA"; that compliance does not exist. Verify AA with axe-core or the browser DevTools accessibility panel.

Verification gate

1. ODIN design gate ≥95% per the <design> block above (tokens / density / paradigm / forbidden list).
2. Anti-slop checklist (§4 Side A AND Side B) returns zero violations.
3. Direction artifact recorded in commit body or PR description: name plus taste anchors plus one-line rationale.
4. Tokens referenced not hardcoded — git grep for raw hex, raw px values in component code; should be empty.
5. Cross-surface invariants honored — spacing subset committed, ≤2 type families, motion named-properties only.
6. Contrast verified via axe-core CLI or DevTools accessibility panel; WCAG 2.1 AA pass on every text surface.
7. Audit tools — data-dense surfaces pass the references/cognitive-load.md checklist (≤1 failure); shipped surfaces have been read against ≥2 of the references/personas.md archetypes.

Anti-patterns

• "Be creative" / temperature-up — generates more slop, not better direction. VS structure or nothing.
• Single-shot palette generation — RLHF over-aligns to purple; the model's first guess is the slop.
• Skipping the direction commit and jumping to component code — components without a direction are coupling without contract.
• Mixing two paradigms on one surface — neo-brutalism shadow on a glass card reads as confusion.
• Ranking directions by implementation effort — short-term cost is the wrong axis; taste fit and depth are the right ones.