visual-design-system

Give a landing page a distinctive, on-brand look — and make the agent actually hold to it.

The failure this fixes

AI generators converge on one identical aesthetic: Inter/Poppins type, a blue-to-purple (indigo/aurora) gradient hero, glassmorphism cards, a centered mx-auto hero with an "eyebrow pill", a three-card feature grid, and the default shadcn fingerprint — "AI slop", the "sea of sameness". It gets worse under iteration: "make it premium" or "be distinctive" ratchets a model toward the average, because it reads "distinctive" as "add more". And prose brand rules are followed only probabilistically and silently dropped under deadline — banning a color in writing has been observed to fail repeatedly while a lint rule fixed it first try. So this skill does two things: choose constrained, distinctive budgets, and encode + enforce them mechanically.

When to use / when NOT to use

Use to choose palette / type / spacing, kill the generic AI look, and set up a machine-readable design system (DESIGN.md, tokens, lint gates) that agents obey.

Not for: choosing the overall look/direction before tokens exist (landing-page-art-direction — it decides, this skill systematizes), the shipped deterministic lint hook (ux-guardrails), the words (landing-page-copywriting), section skeleton (landing-page-structure), animation (scroll-motion), building components (nextjs-landing-page), full WCAG compliance work (landing-page-accessibility — this skill treats contrast only as a floor), or ranking/perf (web-vitals-and-seo).

Workflow

Set the aesthetic budget (hard constraints beat vibes). Palette: ≤4 roles — two neutrals + one ink + one accent, accent reserved for links/primary CTA only. Warm off-white background (#fafaf9), near-black ink (#1c1917) — never pure #fff/#000. Type: 1-2 typefaces, ≥16px body floor. Spacing: 8px scale, ~96px between sections. Show the real product, not stock. Full budgets + the ban-list: references/aesthetic-budgets.md.
Ban the slop tells explicitly. Write down what NOT to do (blue→purple gradients, glassmorphism, sparkle icons, three equal-weight cards, centered eyebrow-pill hero, marquee logo strips) — a Don'ts list built from bad output you have actually seen is the single highest-leverage artifact.
Encode it as DESIGN.md. Machine-readable YAML tokens + prose rationale, tokens defined as roles not raw values, checked into the repo root beside AGENTS.md. Structure, the 8 sections, and the tokens-vs-prose split: references/design-md-and-enforcement.md. DESIGN.md is a fast-converging 2026 convention (Google Labs, alpha), not a ratified standard — verify against the current spec.
Enforce with deterministic gates. Prose is advisory; a failing lint rule is not. Fastest path: install the ux-guardrails hook (this skill set ships it) — write-time spacing/type/contrast/slop checks wired into Claude Code. Full CI stack on top: ESLint no-restricted-imports + custom no-arbitrary-colors; Stylelint declaration-strict-value; TypeScript discriminated unions for variants; visual-regression (Playwright toHaveScreenshot()); axe-core in CI. Cap the agent's self-correction loop and forbid it editing the lint config. Details: references/design-md-and-enforcement.md.

The rules

Constrain the brand hard; keep creative exploration separate. Over-specify the brand tokens (palette, type, spacing) but do not over-specify the creative direction — heavy prompt specification can suppress a model's design strengths and produce more monotonous output. Constrain what must be consistent; leave room to explore what should be distinctive.
Tokens define roles, not values. "primary" is "the main ink color of the page", not a hex. Every consumer references the role, so a rebrand is a one-line edit. Tokens carry what; prose carries when/why. A pure token dump still yields off-brand output.
The Don'ts list is the highest-leverage section — one "don't" per real bad output.
Enforcement beats persuasion. Consistency you enforce survives the agent; consistency you request survives until the agent is in a hurry. The agent never opens Figma — it reads the repo, so the governing spec is tokens-as-code plus a gate it cannot pass without obeying.
Layer the instruction files. Behavioral rules → AGENTS.md/CLAUDE.md; visual spec → a dedicated DESIGN.md loaded only when generating UI (avoids paying the token tax on every turn). Keep root instruction files lean.
Contrast floor (defer depth to accessibility skill): body text 4.5:1, large text and non-text UI 3:1 — a hard gate, not a preference.

Output

A defined palette/type/spacing budget with an explicit ban-list; a DESIGN.md with role-based tokens, prose rationale, and a Don'ts section; and a wired enforcement stack (lint + visual + a11y gates) the agent must pass — verified or version-gated against the current DESIGN.md spec, never citing raw research paths.

References

references/aesthetic-budgets.md - concrete palette/type/spacing budgets, the warm-neutral token set, and the full AI-slop ban-list.
references/design-md-and-enforcement.md - DESIGN.md structure and the 8 sections, tokens-as-roles, authoring methods, Figma MCP/Code Connect, and the deterministic enforcement stack with the fact ledger.