kwakseongjae/oh-my-design / skills/omd-autopilot
omd-autopilot
VerifiedOne-prompt autonomous product design and implementation. Use automatically for broad greenfield UI requests such as 'from scratch', '새 제품/화면을 알아서 만들어줘', or requests that delegate DESIGN.md creation. It decides whether to reuse, establish, refresh, or skip a project design system; asks at most one consequential question batch; then builds and verifies the real surface. Use omd:harness instead only when the user explicitly asks for guided checkpoints.
references/visual-quality-contract.md
View on GitHub# Visual quality contract (patch 7)
Read this in full before PRODUCT_BUILD. These are yes/no gates, not taste.
Distilled from measured competitor analysis (Hallmark slop gates, Pro Max
token layering) and re-anchored to the design-system-first philosophy: the
system is locked once, every page consumes it, and quality means the page is
faithful to the system WITHOUT collapsing into genre defaults.
## 0. Two absolute product rules
- **P0-1 (state switchers).** States are outcomes of real product
interactions only. NEVER render a developer/state switcher in the product
UI — no "Show loading state" radios, no demo toggles, no state menus. If a
brief names states, build the interaction that produces each state.
`data-state` markers live on the real components, never on a switcher.
- **P0-2 (native controls).** NEVER ship native unstyled form controls.
Every radio, checkbox, select, and range is restyled from system tokens
(`appearance: none` + custom mark, or an equivalent overlay) while keeping
the accessible native input functional underneath. A default blue radio on
the page is a shipped bug.
## 1. Token floor (the system must contain these before any page)
- Color in OKLCH, layered: paper / ink / tinted neutrals (5–9 steps, chroma
0.005–0.015) / ONE accent. Pure `#000`/`#fff` are not base colors. Every
component reads `var(--color-*)`; an inline hex outside the token file is
a gate failure.
- Space: 4pt scale with ROLE names, and it must extend past control padding
into section air — include steps around 2.5rem / 4rem / 6rem and a
viewport gutter `clamp()`. Section transitions use those; if every gap on
the page is the same step, the rhythm failed.
- Type: a display/body pair (plus at most one outlier face), even from local
stacks — a single `system-ui` line has no genre. Display step: `clamp()`
to ≥2× body size, tracking −0.02 to −0.045em, line-height 1.0–1.1,
weight ≥700. A LABEL role exists (smaller, uppercase, wide tracking,
muted). Font-family count ≤3; no italic display; body line-height ~1.5,
prose measure 45–75ch.
- Motion tokens per the motion section of the skill; additionally banned:
`transition: all`, bounce/overshoot easing on UI state, more than one
hover effect stack (translate+scale+shadow+color together), animating
layout properties, focus rings that fade in.
## 2. Surface genre (pick one, commit)
Choose ONE surface genre per product and hold it on every page:
hairline-rule bands / no-radius ink plates / shadowed tiles. Never repeat
the same `radius + 1px border` box on cards, banners, dialogs, and controls
alike — that is a widget kit, not a design. Elevation states are tokens
(rest / hover / selected), not improvised shadows.
Accent is a SIGNAL, not an area: a dot, a 2–3px rule, a selection ring, or
one inverted chip — total accent coverage ≲5% of any viewport. Never both a
full-card selected wash and a full accent CTA fill on the same screen.
## 3. Component form
- Interactive components implement the full state matrix in code:
default / hover / focus-visible / active / disabled (+ loading / error /
success where they exist). Hover must change surface (background, border,
or shadow), not merely translate. Selection gets a PERSISTENT mark
distinct from hover (ring, mark, rule). 3D tilt only inside
`@media (hover: hover) and (pointer: fine)`.
- Input geometry: border-width never changes between states; focus is
`outline` (reserved slot, instant); input height equals adjacent button
height; helper/error slots keep `min-height: 1lh`; disabled = opacity +
`cursor: not-allowed` + native `disabled` (all three).
- Cards separate roles: badge / short proper name / label row / action
affordance. Empty states and dialogs are DISTINCT surfaces (dashed border
or native `<dialog>` with styled `::backdrop`), never a muted paragraph
inside the same card.
- Icons: one family, one stroke width; never emoji as feature icons; never
two icon libraries. Visuals tier: typography → CSS art → inline SVG →
(last) libraries. Nothing decorative without `aria-hidden`.
## 4. Page macro-structure
- The first screen is composed, not centered-stacked: provided imagery goes
into a copy/figure diptych grid (never a cropped strip above a thin
header bar); without imagery, open with a short display statement
(~12–22ch). Fold discipline at 1280×800: eyebrow/title/lede/primary
action visible; hero bottom padding ≥1.3× top.
- Content sites never get the hamburger + sidebar app shell. Sections
transition through distinct devices (rule, plate inversion, background
shift), not identical whitespace.
- Page chrome exists: a wordmark in the display role, a sample/issue line
where honesty requires it, and a footer/colophon — the page must not
start and end with bare widgets. But nav/footer are CANONICAL: designed
once in the system, identical on every page, current page marked.
- Fixed-count galleries use authored 1/2/3-column breakpoints, never
`auto-fit` (the stretched last row is a failure). Image grid tracks are
`minmax(0, 1fr)`; display headings get `overflow-wrap: anywhere;
min-width: 0`. No horizontal scroll 320–1920 (`overflow-x: clip`, never
`hidden`). No nav/CTA/tab label wraps to two lines at any width.
- Contrast: body ≥4.5:1, large text/icons/focus ≥3:1; accent fills define
an ink counterpart token; dark sections flip text tokens explicitly.
## 5. Multi-page consistency (the system's whole point)
- ONE shared stylesheet holds tokens and components; pages add only page
layout. `:root` tokens are defined exactly once — a re-declared token
with a different value on another page is a critical failure.
- Body, headings, nav, and primary action must compute to identical styles
on every page. Every internal link resolves. The nav link set is
identical everywhere with `aria-current` on the active page.
## 5a. Locale typography (the content language picks the rules)
Detect the content language and apply its typography contract; Latin rules
do NOT transfer.
Korean (ko):
- Font stacks must COVER Hangul for every role. Body and display:
`"Pretendard Variable", Pretendard, "Apple SD Gothic Neo", "Noto Sans KR",
"Malgun Gothic", sans-serif`. NEVER let a Latin serif/display face fall
back to a system serif for Hangul — if no Korean display face exists
locally, build display hierarchy with WEIGHT (700–800) and size, same
gothic family.
- `word-break: keep-all` on all Korean prose and headings;
`overflow-wrap: anywhere` stays on display headings.
- Body line-height 1.6–1.8 (Hangul needs more than Latin). Display 1.15–1.3.
- NO tight letter-spacing on Hangul (−0.01em is the floor; the −0.03em Latin
display rule does not apply). Numbers/latin fragments inside Korean text
may use `font-feature-settings "tnum"` and tighter tracking.
- Prices and counts: thousands separators, unit after the number (129,000원).
## 5b. UI copy register (user language only)
- Implementation vocabulary NEVER appears in the UI: no file names
(data.json), no field/function names (products.length, review_count), no
dataset/framework words. Screens speak the user's language.
- Aggregate definitions are stated in USER terms: "리뷰 많은 순 상위 4개",
never "data.json을 내림차순 정렬한 상위 4건".
- Sample/fiction disclosure is ONE quiet line in the footer (plus detail
pages where money is shown). Never a top banner, never repeated per
section.
## 5c. SPA fundamentals (route changes behave like pages)
- Every route change scrolls to top (and restores position on back/forward);
a detail page must never open mid-scroll.
- Focus moves to the new page's h1 (or main) on route change.
- `document.title` updates per page ("상품명 — 서비스명").
- Below-the-fold images `loading="lazy"`; the LCP image is not.
## 5d. Component anatomy (fill the anatomy, don't improvise a card)
- Card anatomy is a spec, not a suggestion: media block and text block share
the SAME width; body padding is one token step; the internal rows
(label → title → price → meta → action) use a fixed gap scale
(e.g. 4/8/12) — never a uniform gap between all rows, price sits closest
to the title, action is separated by the largest step or a rule.
- List rows: one baseline grid; icons/badges vertically centered to the
first text line; consistent right-alignment for numbers.
- Never mix aligned and unaligned edges in one card (image edge = text
edge = action edge).
## 5e. Genre–domain fit (the surface genre must match the product)
- Commerce/catalog: card surfaces with clear boundaries (shadowed tile or
bordered card), price hierarchy dominant, generous media. Hairline
editorial rules are for reading surfaces, not shop grids.
- Ops/dashboard: density first — tables, compact rows, status tokens.
- Editorial/content: rule bands, plates, long-form measure.
- Choose the genre FROM the domain, then commit (rule in §2 still holds).
## 6. Self-critique loop (mandatory, one round)
Before declaring the product finished, run ONE fast critique pass and write
`critique.md` in the run directory:
1. Score 5 axes 1–5: Philosophy (does the page take a stance) / Hierarchy
(1-2-3 order readable in 2s) / Execution (rules, accent, focus, contrast
are spec-true) / Specificity (looks like THIS brief, not any product) /
**System Fidelity** (every value traces to a locked token AND the page
has not collapsed into genre defaults: 3-column icon cards, centered
hero stack, AI nav/footer, purple gradients).
2. Sweep `references/slop-gates.md` IN FULL (numbered, lossless — do not
rely on this contract's summaries alone); list every hit by number.
3. Any axis <3 or any P0/critical gate hit → fix now, once. Record
before/after in `critique.md`. Never fake the scores — the audit that
follows reads the same DOM.
The SELF-WALK (journey verbs) still runs; this critique is about form, the
SELF-WALK about function. Both fit inside the product budget because they
are checklists, not rebuilds.
## 7. Render-feedback pass (two-pass quality)
Source-level self-review cannot see typography fallbacks, spacing rhythm,
or scroll behavior — only a rendered screen can. When an external reviewer
supplies RENDERED-SCREENSHOT findings for your build (a numbered punch
list), treat every item as a confirmed defect observed on the real screen:
fix each one surgically inside the existing system (token/anatomy-level
changes, not re-architecture), keep all prior contracts intact, and reply
by listing fix-by-fix what changed. Do not dispute a finding you cannot
see; the reviewer is looking at the pixels.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| brainstormingobra/superpowersYou MUST use this before any creative work - creating features, building components, adding functionality, modifying behavior, or planning anything new, in software or out of it (a talk, a business, a renovation). | Stars: 323k | Updated | Claude CodeCodex | Verified | Install |
| brand-guidelinesanthropics/skillsApplies Anthropic's official brand colors and typography to any sort of artifact that may benefit from having Anthropic's look-and-feel. Use it when brand colors or style guidelines, visual formatting, or company design standards apply. | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| frontend-designanthropics/skillsGuidance for distinctive, intentional visual design when building new UI or reshaping an existing one. Helps with aesthetic direction, typography, and making choices that don't read as templated defaults. | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| mcp-builderanthropics/skillsGuide for creating high-quality MCP (Model Context Protocol) servers that enable LLMs to interact with external services through well-designed tools. Use when building MCP servers to integrate external APIs or services, whether in Python (FastMCP) or Node/TypeScript (MCP SDK). | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |