Use when designing or restyling a Hwaro site — choosing a visual direction, customizing templates and CSS, building or retheming a theme, picking typography and color, or improving an existing layout. Read the brief, declare a Design Read and set the three dials, interview only when intent is genuinely ambiguous, then produce distinctive, production-grade design within Hwaro's Crinja template + CSS-variable token system — under a strict anti-slop discipline and a mechanical pre-flight check. Pair with the `hwaro` skill, which covers operating the CLI itself.
설치
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
Use when designing or restyling a Hwaro site — choosing a visual direction, customizing templates and CSS, building or retheming a theme, picking typography and color, or improving an existing layout. Read the brief, declare a Design Read and set the three dials, interview only when intent is genuinely ambiguous, then produce distinctive, production-grade design within Hwaro's Crinja template + CSS-variable token system — under a strict anti-slop discipline and a mechanical pre-flight check. Pair with the `hwaro` skill, which covers operating the CLI itself.
Hwaro Design
Your job is to design Hwaro sites that look intentional, distinctive, and
genuinely good — and that reflect this user's taste, not a generic default.
A static site generator makes the plumbing trivial; the design is where the site
either feels considered or feels like every other template. Aim for the former.
This skill assumes the project is a Hwaro site (config.toml + templates/ +
static/). For scaffolding, building, and serving, use the hwaro skill;
iterate with hwaro serve (live reload) so the user sees changes instantly.
Every rule below is contextual — none fires automatically. First read the
brief, then pull only what fits. But the bans in Section 9 and the Pre-Flight
Check in Section 12 are hard filters: output that violates them is not done.
0. The Design Read — read the room before anything else
Before touching code or setting dials, infer what the user actually wants.
Most AI design output is bad because the model jumps to a default aesthetic
instead of reading the room.
0.A Read these signals first
Site kind — blog / docs / portfolio (dev, designer, studio) / landing
(product, event, agency) / book / shop-adjacent brochure / redesign
(preserve vs. overhaul).
Vibe words the user used — "minimalist", "calm", "Linear-style",
"Awwwards", "brutalist", "premium consumer", "Apple-y", "playful",
"serious B2B", "editorial", "glassy", "dark tech", "warm-handcrafted".
Reference signals — URLs they linked, screenshots they pasted, products
or brands they named.
Audience — recruiters scanning a portfolio vs. developers reading docs
vs. design-conscious consumers. The audience picks the aesthetic, not your
taste.
Brand assets that already exist — logo, color, type, photography. For
redesigns these are starting material, not optional input (Section 10).
Quiet constraints — accessibility-first audiences, public-sector,
regulated industries, kids' content, CJK-primary readership (keep the CJK
font fallbacks — see Section 4.1). These constraints OVERRIDE aesthetic
preference.
0.B Declare a one-line Design Read before generating
Before any code, state in one line: "Reading this as: <site kind> for
<audience>, with a <vibe> language, leaning toward <aesthetic family>" —
followed by the three dial values (Section 1) and why.
Example reads:
"Reading this as: developer blog for peers, with a calm technical-editorial
language, leaning toward mono-accented minimalism. Dials: 5 / 3 / 4."
"Reading this as: product landing for design-conscious indie devs, with a
confident dark-tech language, leaning toward near-black canvas + single
electric accent. Dials: 7 / 6 / 3."
"Reading this as: redesign-preserve of an existing docs site — extract the
current tokens first, evolve typography and rhythm, keep the IA. Dials: match
existing."
0.C Question policy — hybrid
If you can confidently infer the direction from the brief, conversation, or
the existing site: do not ask. Declare the Design Read and proceed. A
confident direction the user can react to beats an interview they didn't want.
If the read genuinely diverges (e.g. "Linear-clean or
Awwwards-experimental?" would produce different sites), ask a short round —
1–2 focused questions, not a survey. Use AskUserQuestion when available:
2–4 concrete options per question with short descriptions, and lean on its
option previews — a token-value block or a small ASCII layout sketch per
option beats adjectives; people pick a palette they can see.
Run a full interview only when the user asks to explore taste ("interview
me", "help me figure out what I want") or a greenfield brief gives you nothing
to read. Then cover, in order, stopping once you can commit: purpose &
audience · personality + 1–3 reference sites · color mood incl. light/dark ·
type feel · density & layout · motion appetite · constraints.
Either way, for any substantial engagement play the brief back in 4–6
lines ("Warm editorial blog: cream paper background, single rust accent,
Charter-style serif headings over a clean sans body, generous whitespace,
motion limited to gentle link transitions") and confirm before building big.
The brief is the contract you design against.
If a confirmed brief already exists earlier in the conversation, don't
re-interview — proceed from it, and only revisit taste if the user's
direction changed.
0.D Anti-Default Discipline
Do not default to: AI-purple gradients, centered hero over dark mesh, three
equal feature cards with an emoji each, generic glassmorphism on everything,
infinite-loop micro-animations, system-font-with-no-personality — and, the
Hwaro-specific one, shipping the ember scaffold palette unchanged and calling
it a design. The ember tokens are a starting vocabulary, not the answer to
a brief. Reach past the defaults deliberately, based on the Design Read.
1. The Three Dials
After the Design Read, set three dials. Every layout, motion, and density
decision below is gated by these.
VISUAL_DENSITY — 1 = art gallery / airy, 10 = cockpit / packed data
State the values in the Design Read. Never silently assume a baseline —
reason them from the brief. Overrides happen conversationally ("crank the
motion to 8").
MOBILE OVERRIDE: for levels 4–10, asymmetric layouts MUST collapse to a
strict single column below 768px. Declare the collapse in the same rule
block, not "it'll probably reflow."
8–10 (Choreography): scroll-driven reveals via CSS
animation-timeline: view() (behind @supports) or an IntersectionObserver
(Section 7). window.addEventListener("scroll", …) is a hard ban, not a
prefer-not.
4–7 (Daily site): standard rhythm (4–6rem section padding) — the
--space-* scale as-shipped.
8–10 (Cockpit): tight paddings, no card boxes, hairlines separate data,
--font-mono for all numbers.
2. How design lives in a Hwaro site
Know the actual surfaces you can touch — this is what makes the design real
and not just advice.
Templates (Crinja / Jinja2-compatible) — the structure
Under templates/. Hwaro supports two composition styles and the built-in
scaffolds use both, so open the project's templates/ and see which before
editing: {% extends "base.html" %} inheritance filling {% block content %}
(the docs scaffold works this way), and {% include "header.html" %} /
{% include "footer.html" %} partial composition (blog, book, simple).
Other files you'll meet: page.html / section.html (content layouts),
index.html (home), taxonomy*.html, and partials like nav.html /
sidebar.html. Layout changes (what wraps what) happen here; look (color,
type, spacing) happens in CSS. Keep markup semantic and class-driven so the CSS
can do the work.
Internal URLs must carry the base URL. The scaffolds write every
site-internal href/src as {{ base_url }}{{ page.url }},
{{ base_url }}{{ lang_prefix }}/about/, {{ base_url }}/css/style.css —
never a bare /about/. Keep that pattern in every template you author, or the
site 404s the moment it's deployed under a subpath (GitHub Pages project
sites).
CSS delivery — three options, pick deliberately
Inlined <style> — the simple scaffold ships its CSS in a <style>
block inside header.html: zero extra requests, great for small sites; edit
that block directly. (blog / book / docs instead ship an external
static/css/style.css — option 2; bare ships no CSS at all, so you bring
your own.)
Static stylesheet + [auto_includes] — drop files in static/assets/css/
(and …/js) and emit {{ auto_includes_css }} / {{ auto_includes_js }} in
your head/body. (These output raw HTML; Hwaro disables template autoescape, so
the scaffolds write them without | safe — the docs add | safe as a
harmless, portable convention you can keep.) Good when CSS grows past "one block."
[assets] pipeline — declare bundles in config.toml, reference with
{{ asset(name='main.css') }}; you get concatenation, minification, and
content-hash fingerprinting (cache busting) for free.
There is a built-in SCSS compiler (opt-in via [sass] enabled = true),
but no Tailwind step and no component framework. The built-in compiler
covers the practical subset — variables, nesting/&, partials with
@use/@import, mixins with @content, interpolation — but NOT control
flow (@if/@each), arithmetic, or built-in functions (lighten() etc.),
so framework-grade SCSS won't compile. Default to modern plain CSS (custom
properties, nesting where supported, color-mix(), clamp()); use [sass]
when the brief calls for it, or run Tailwind/PostCSS/full dart-sass via
build hooks and point Hwaro at the compiled output. Don't assume an npm
design system exists. When a brief names an aesthetic
(glass, bento, brutalist, editorial, dark tech, aurora), that's a CSS
language you build honestly by hand — there is no official package for any of
them, on any stack.
Design tokens are your single source of truth
Every Hwaro scaffold themes itself through CSS custom properties in :root.
This is the right pattern: define the system once, theme by editing tokens, and
get dark mode + restyles almost for free. All eight scaffolds share one
"Hwaro Ember" vocabulary: every color token is a light-dark(light, dark) pair
resolved by color-scheme, so every scaffold (including simple) follows the
reader's OS scheme automatically. The vocabulary is (values exact; font stacks
abbreviated — the real ones are longer and include CJK fallbacks like
"Noto Serif KR", so don't strip those if your audience needs CJK):
Every scaffold shares this exact vocabulary. It is emitted from one place —
DesignTokens.root_block in src/services/scaffolds/design_tokens.cr in the
Hwaro repo — so the old per-scaffold name drift is gone (--primary-hover,
--bg-secondary, and --border-light were renamed to --primary-strong,
--bg-subtle, and --border-subtle). The only per-scaffold additions are
layout tokens (--header-h, --sidebar-w, --content-max-w, --bg-sidebar,
radius overrides) injected into the same :root. A
@supports not (color: light-dark(#000, #fff)) block pins the static light
palette for pre-2024 browsers, so a *-dark site renders light there —
readable and on-brand rather than broken.
Retheme by rewriting token values, not by adding parallel systems. The type
scale (--step--1…--step-4), spacing rhythm (--space-1…--space-8),
--measure, radii, shadows, and --transition already exist — don't re-create
them under new names; override their values (a different ratio, a denser rhythm)
and build every component against them. Never hardcode a raw color in a rule:
reach for the existing pair, or color-mix() off one.
/* A custom retheme = override the light-dark() pairs in :root.
Always set BOTH sides — a light-only override ships a broken dark scheme. */:root {
--primary: light-dark(#3a6ea5, #7fb2e5);
--primary-strong: light-dark(#2c567f, #a3c9ee);
--bg: light-dark(#f7f9fb, #101418);
/* …continue through the ink ramp, surfaces, and --code-* slots… */
}
/* Manual user toggle (scaffolds do NOT ship one): flip color-scheme via a
data attribute + a small JS toggle that sets it. */:root[data-theme="dark"] { color-scheme: dark; }
:root[data-theme="light"] { color-scheme: light; }
Reality check: dark mode is automatic now — don't re-plumb it.
Every scaffold is light and dark out of the box: all color tokens are
light-dark() pairs under color-scheme: light dark, so the site follows
the OS scheme with zero extra CSS.
Forcing dark permanently = append :root { color-scheme: dark; } as the
last rule of the sheet — that is literally all the *-dark scaffolds do.
(Delete that rule to restore automatic switching.)
A custom retheme = override the light-dark() pairs in :root — always
supply both sides of each pair, or one scheme ships broken.
A user-facing toggle = the [data-theme] pattern above plus a small JS
toggle; you write the toggle, the tokens do the rest.
A hygiene spec (spec/unit/scaffold_token_hygiene_spec.cr) enforces that
scaffold CSS has no hardcoded colors outside the token definitions —
when you edit scaffold source CSS, keep it green by routing every color
through a token.
Code blocks are a design surface
Hwaro colors code server-side (Tartrazine), and the scaffolds' .hljs-*
rules read entirely from the nine --code-* tokens — so recoloring syntax
means editing those light-dark() pairs in :root, not touching the .hljs-*
rules themselves. Tie them to your palette (e.g. keywords near your accent) for
cohesion; both schemes come along automatically.
3. The quality bar
Production-grade design is a stack of deliberate decisions. Hit every layer.
A point of view. Pick one organizing idea (Swiss/grid precision, warm
editorial, brutalist mono, retro terminal, glassy modern, dark tech…) and let
it govern every choice. Coherence reads as quality.
Typography does the heavy lifting. A deliberate pairing (or one great
family with real weight range), a modular scale, tuned line-height and
measure. Details in Section 4.1.
Color with intent. One dominant surface family, considered neutrals, ONE
confident accent. Details in Section 4.2.
A spatial system. One spacing scale used everywhere. Consistent rhythm,
alignment, and generous negative space are what make a layout feel composed.
Clear hierarchy & a focal point. Every screen has an obvious entry point
and an unambiguous reading order — established by size, weight, color, and
space, not boxes everywhere.
Texture & detail, used with restraint. Hairline borders, a single
considered shadow elevation, a subtle gradient or grain, one small recurring
motif (the scaffolds use one short accent rule under the page title). Detail
signals care; clutter signals the opposite.
Motion with purpose. Section 7. Always honor prefers-reduced-motion.
Responsive by construction. Design the small screen and the large screen,
not just a desktop that shrinks. Fluid type/space via clamp(), real touch
targets.
Accessibility is part of "good," not a checkbox. Semantic HTML, visible
:focus-visible, a skip link, alt text, sufficient contrast, logical heading
order.
LLMs default to clichés. Override these defaults proactively. Each rule has a
context-aware override path.
4.1 Typography
Display / headlines: large steps of the scale (--step-3/--step-4),
letter-spacing: -0.02em-ish tightening, line-height near 1–1.15,
text-wrap: balance.
Body:--step-0, line-height 1.5–1.7, max-width: var(--measure)
(60–75ch), text-wrap: pretty on paragraphs.
Fonts are self-hosted.@font-face + font-display: swap, subset files,
limit families/weights — fonts are usually the biggest perf cost of a
"designed" site. Never <link> Google Fonts in production. The scaffolds
embed Charter via OFL Charis SIL so headings render the same off-Apple.
Keep the CJK fallbacks. The shipped stacks include "Noto Serif KR"-class
fallbacks; if the audience reads CJK, stripping them breaks the design where
it matters most. (This also applies to OG images — CJK-capable font chains.)
SERIF DISCIPLINE. The ember scaffolds ship Charter serif headings, and
for content sites (blog, docs, book — genuinely editorial surfaces) that
default is legitimate. But when the brief is a landing page, portfolio,
agency, or product site, serif is very discouraged as the default
reach. "It feels creative / premium / editorial" is NOT a reason. Serif is
acceptable only when the brand brief literally names a serif, or the
aesthetic family is genuinely editorial / luxury / publication / heritage AND
you can articulate why this serif fits this brand. Otherwise default to a
characterful sans display. Specifically banned as default reaches:
Fraunces and Instrument Serif (the two LLM-favorite display serifs). If a
serif is justified, don't reuse the same one across consecutive projects.
EMPHASIS RULE. To emphasize a word inside a headline, use italic or
bold of the SAME family. Do not inject a serif word into a sans headline
(or vice versa) for visual interest — mixed-family emphasis is amateur.
ITALIC DESCENDER CLEARANCE. Italic display words containing y g j p q
get clipped at line-height: 1. Use ≥ 1.1 and reserve a few px of
padding-bottom on the wrapper. Audit every italic display word before
shipping.
No oversized H1s that just scream. Control hierarchy with weight, color,
and space, not raw scale alone.
4.2 Color calibration
Max 1 accent color, saturation < 80% by default. Neutrals are never pure
#000/#fff (the ember ink ramp is the model). Use color-mix() for
tints/states instead of inventing new hex values.
THE LILA RULE. The "AI purple/blue glow" aesthetic is banned as a
default. No automatic purple button glows, no random neon gradients. Neutral
bases with one high-contrast singular accent (emerald, electric blue, deep
rose, burnt orange…). Override: if the brand explicitly is purple, embrace
it — with a consistent palette and restrained gradients, not gradient slop.
COLOR CONSISTENCY LOCK. Once the accent is chosen, it is THE accent for
the whole site. A warm-gray site does not get a blue CTA in the footer; a
rose-accented site does not get a teal badge. In Hwaro terms: everything
reads from --primary / --primary-strong — if you're tempted to add a
second accent variable, you're breaking the lock.
One palette per site. Don't fluctuate between warm and cool grays.
PREMIUM-CONSUMER PALETTE BAN. For premium-consumer briefs (cookware,
wellness, artisan goods, luxury, DTC home goods) the LLM default is warm
cream/beige surfaces + brass/clay/oxblood accents + espresso near-black text.
Note: the ember scaffold palette IS this family (warm paper #faf7f2,
rust accent, warm ink). That's fine as Hwaro's out-of-the-box identity — but
reaching for it (or just keeping it) as the design for a premium-consumer
client brief is the lazy move. Rotate to a genuinely different family:
Cold luxury: silver-gray + chrome + smoke
Forest: deep green + bone + amber accent
Black and tan: true off-black + warm tan, sharp contrast, no beige
Cobalt + cream: saturated blue against a single neutral
Terracotta + slate: warm rust against cool gray
Monochrome + one pop: off-white + off-black + one bright accent
Override: warm-craft beige is acceptable when the brand brief explicitly
names those colors or the identity is genuinely vintage/artisan AND you can
say why. Defaulting to it because "this is a cookware brief" is banned.
Both sides, always. Every custom color is a light-dark() pair with both
sides supplied (Section 2). A light-only override ships a broken dark scheme.
4.3 Layout diversification
ANTI-CENTER BIAS. Centered hero / H1 sections are avoided when
DESIGN_VARIANCE > 4. Prefer split (50/50), left-aligned content with
right-aligned asset, asymmetric whitespace, or a pinned structure.
Override: centered is fine for editorial / manifesto briefs where the
message itself is the design — and for docs/book content pages, which are
reading surfaces, not compositions.
4.4 Materiality, shadows, cards
Use cards ONLY when elevation communicates real hierarchy. Otherwise group
with a top border, row dividers used sparsely, or negative space.
When a shadow is used, tint it toward the background hue (the ember
--shadow* tokens already do this). No pure-black drop shadows on light
backgrounds.
For VISUAL_DENSITY > 7: generic card containers are banned; data breathes
in plain layout with hairlines.
SHAPE CONSISTENCY LOCK. One corner-radius system per site — that's what
--radius / --radius-sm are for. All-sharp (0), all-soft (12–16px), or
all-pill for interactive — pick one, or document the mixed rule ("buttons
pill, cards 16px, inputs 8px") and follow it everywhere. Round buttons in a
square layout is broken design.
4.5 Interactive states
Tactile feedback: on :active, a 1px translate or scale(0.98) to
simulate a physical push. Short eased transitions on all interactive
elements (var(--transition)).
Visible :focus-visible on every interactive element. A skip link at the
top of base.html/header.html.
BUTTON CONTRAST CHECK. Every CTA's text is readable against its own
background: WCAG AA (4.5:1 body, 3:1 large/UI). White-on-white, ghost
buttons over photos with no scrim/stroke — banned. Audit every CTA.
CTA BUTTON WRAP BAN. Button text fits on one line at desktop. If a label
wraps, shorten the label (1–3 words for primary CTAs) or widen the button.
NO DUPLICATE CTA INTENT. "Get in touch" + "Contact us" + "Let's talk" on
one page = one intent, three labels = fail. Pick ONE label per intent and use
it everywhere (nav, hero, footer). Same for "View work" / "Browse projects".
FORM CONTRAST CHECK. Inputs, placeholders, focus rings, helper and error
text all pass WCAG AA against the section background. Label ABOVE input;
error text below. No placeholder-as-label. Ever.
4.6 Layout discipline (hard rules — failing any of these is shipping broken work)
Hero MUST fit the initial viewport. Headline ≤ 2 lines on desktop,
subtext ≤ 20 words AND ≤ 3–4 lines, CTA visible without scroll. If the copy
is too long, cut copy or reduce scale — never let the hero overflow.
Use min-height: 100dvh for full-height heroes, never 100vh (iOS address
bar jump).
Hero font-scale discipline. Plan font size and asset size together.
--step-4 territory only when the headline is 3–5 words; a 4-line hero
headline is always a font-size error, never a copy-length error.
HERO TOP PADDING CAP. Max ~6rem top padding at desktop. If the hero needs
more breathing room, increase font scale or asset size, not top padding.
HERO STACK DISCIPLINE (max 4 text elements): eyebrow OR brand strip (or
neither) · headline · subtext · CTAs (1 primary + max 1 secondary). BANNED in
the hero: tiny tagline below CTAs, trust micro-strip, pricing teaser, feature
bullet list, social-proof avatar row — those move to sections below.
"Trusted by" logo wall belongs UNDER the hero, never inside it.
Navigation renders on ONE line at desktop, height ≤ 80px (default
64–72px, the scaffolds' --header-h). If items don't fit at 1024px, condense
labels or move to a menu. A two-line desktop nav is broken.
Bento grids have rhythm and exact cell count. N items → N cells; no blank
filler tiles. Vary composition; and at least 2–3 cells in any multi-cell grid
need real visual variation (an image, a brand-appropriate gradient, a
pattern, a tinted background) — an all-same-surface bento with only
typography inside reads as AI default.
SECTION-LAYOUT-REPETITION BAN. One layout family (3-col cards, full-width
quote, split text+image…) appears at most ONCE per page. A landing page with
8 sections uses at least 4 different layout families.
ZIGZAG ALTERNATION CAP. Max 2 consecutive image+text split sections. The
3rd consecutive zigzag is a pre-flight fail — break it with a full-width
section, a vertical stack, a bento, or a marquee.
EYEBROW RESTRAINT (the #1 violated rule). An eyebrow is the small
uppercase wide-tracking label above a section headline (CSS signature:
text-transform: uppercase; letter-spacing: 0.18em; font-size ~11px). Hard
rule: max 1 eyebrow per 3 sections, hero counts as one; if section A has
one, the next two don't. The check is mechanical: count
uppercase-tracked micro-labels across templates; count > ⌈sections / 3⌉
fails. Instead of an eyebrow: drop it — the headline alone is enough.
SPLIT-HEADER BAN. "Left big headline + right small floating explainer
paragraph" as a section header is banned as default. Stack vertically
(headline, then body at --measure). Reach for a split header only when the
right column carries a real visual or interactive element.
Mobile collapse is explicit per section. Every multi-column layout
declares its < 768px fallback in the same stylesheet section. No "the grid
will probably handle it."
4.7 Content density & copy
Landing pages live on the first impression, not the full read. Cut
ruthlessly.
Default section shape: short headline (≤ 8 words) + short sub-paragraph
(≤ 25 words) + one visual asset OR one CTA. More must be justified by the
section's job. (Docs and book content pages are exempt — they're reading
surfaces; this governs marketing/landing/home compositions.)
No data-dump sections. A 20-row table or 30-item list on a marketing page
is the wrong layout: top 3–5 highlights + "view full list" link, a marquee /
scroll-snap row for breadth, or a separate page if the data is the product.
Long lists need a different component, not a longer list. > 5 items:
2-column grouped split, card grid, tabs/accordion, horizontal scroll-snap
pills, or a marquee. A 10-row spec sheet with a hairline under every row is
the worst default — group into 2–3 chunks with sparse dividers, or promote
3–4 hero specs to display tiles and collapse the rest behind a disclosure.
COPY SELF-AUDIT (mandatory before ship). Re-read every visible string
(headlines, eyebrows, buttons, captions, alt text, footer). Flag and rewrite
anything grammatically broken, with unclear referents, or that reads like an
LLM trying to sound thoughtful (forced metaphors, mock-poetic micro-meta,
fake-craftsman labels). If unsure whether a string makes sense, replace it
with a plain functional sentence. AI-cute copy is worse than boring copy.
Fake-precise numbers are flagged.92%, 4.1×, 5.8 mm either come
from real data, are explicitly labeled as sample data, or are banned. Don't
fake engineering precision the brand doesn't claim.
One copy register per page. Don't mix technical-mono metadata, editorial
prose, and marketing punch unless the brand voice explicitly calls for it.
Quotes & testimonials: max 3 lines of quote body — a landing-page quote
is a snippet, not the review. Attribution is name + role (+ company), never
name alone. Real typographic quotes ("") or none — not straight ASCII.
4.8 Image & visual asset strategy
Landing pages and portfolios are visual products. Text-only pages with
fake-screenshot divs are slop.
Priority order for visual assets:
Image-generation tool first. If ANY image-gen tool is available in the
environment, use it to create section-specific assets: hero photography,
product shots, texture backgrounds — at the right aspect ratio per section.
Save the results into the page bundle or static/images/ so they're real
files the build owns.
Placeholder photography second. No gen tool →
https://picsum.photos/seed/{descriptive-seed}/{w}/{h} (seed describes the
section), or stock/brand URLs the brief provides. Treat hotlinks as
placeholders: for the shipped site, download assets into the project so
the static build is self-contained.
Last resort: tell the user. Leave clearly-labeled placeholder slots
(<!-- TODO: hero product photo, 1600x1200 -->) and end with: "This page
needs real images at: [placements]. Please generate or provide them." Do
NOT fill the page with hand-rolled SVG illustrations or div-based fakes.
Plumbing (Hwaro): run every real raster asset through
resize_image(path=…, width=…) — it returns .url, plus .lqip (blur-up
placeholder) and .dominant_color when LQIP is enabled — and build proper
srcset/sizes with lazy-loading. Don't ship one giant image. The built-in
processor resizes JPEG/PNG (BMP) only; for WebP/AVIF or aggressive optimization
run a build hook — resize_image passes unsupported formats through unresized.
Rules:
Even minimalist sites need real images. A pure-text landing page is not
minimalism, it's incomplete. Even a restrained editorial site needs 2–3 real
images (hero, one product/lifestyle shot, one supporting). Blog/docs content
pages are exempt; their home/landing compositions are not.
Div-based fake screenshots are banned. No fake task lists, fake
terminals, fake dashboards built from styled <div>s. Use a real screenshot,
a generated image, or skip the preview.
Hero needs a real visual. Text + gradient blob is a placeholder, not a
hero. (Exception: genuine editorial-manifesto heroes where type IS the
design.)
Real logos for social proof. A "Trusted by" wall uses real SVG marks —
vendor them into static/ (e.g. from Simple Icons) rather than hotlinking a
CDN; ensure they render in both schemes (single-color via currentColor or a
token). For invented brands, draw a simple monogram mark — a plain text
wordmark row looks generic. Logo wall = logos only: no industry/category
captions under each logo.
4.9 Icons & emoji
Use an established icon set, one family per site. Vendor the SVGs you
need (Phosphor, Tabler, Radix, Lucide et al. all offer per-icon SVG
downloads) into static/icons/ or inline them as partials. Standardize
stroke-width across the set. Never hand-draw icon paths from scratch.
Hand-rolled decorative SVGs (custom illustrations, logos) are strongly
discouraged as a default — acceptable only for a single simple geometric
mark, or when the user explicitly asks.
Emoji are discouraged in markup and visible text — the emoji-per-card
feature row is a top AI tell. Use icon glyphs. Override: an explicitly
playful/chat-style brief, sparingly, with intent.
5. Signature detail
Give the design ONE small recurring mark — an accent rule (the scaffolds use a
short ember rule under the page title), a marker bullet, a consistent hover
behavior — rather than decorating everything. One motif, applied consistently,
reads as identity; five motifs read as noise.
6. Pattern vocabulary (names to design with)
A vocabulary, not a library — know the names so you can propose and reason
about them. Implementation cost on a static site matters: prefer pure CSS,
then small vanilla JS, and treat heavy JS as a deliberate, justified
exception (there is no bundler unless you add one via build hooks).
Hero paradigms: asymmetric split hero · editorial manifesto hero (type as
poster) · media-mask hero (type cut out over image/video) · kinetic-type
hero · scroll-pinned hero (CSS position: sticky).
Scroll effects: sticky scroll stack · zoom parallax · scroll progress
bar (animation-timeline: scroll()) · reveal-on-enter
(animation-timeline: view() or IntersectionObserver).
Typography & text: kinetic marquee (CSS keyframes; max one per page) ·
text mask reveal · circular text path (SVG <textPath>) · gradient stroke
text.
Micro-interactions: directional hover fill · ripple click · skeleton
shimmer · animated SVG line drawing (stroke-dasharray).
Heavy-JS patterns (scroll hijack, WebGL scenes, physics cursors) are usually
the wrong trade for a static site; if the brief truly demands one, isolate it
in one small script, lazy-load it, and degrade gracefully without it.
7. Motion for a static site
No React, no animation framework — and none needed. The toolkit, in order:
CSS transitions for all interactive states (var(--transition),
cubic-bezier(0.16, 1, 0.3, 1) for bigger moves).
CSS keyframes + animation-delay cascades for load-in staggers
(animation-delay: calc(var(--i) * 60ms) with a per-item --i).
CSS scroll-driven animations (animation-timeline: view() /
scroll()) for reveal and progress effects — always behind
@supports (animation-timeline: view()) so unsupported browsers get static
content, not broken content.
IntersectionObserver (small vanilla script) as the portable reveal
mechanism:
<script>const io = newIntersectionObserver((entries) => {
for (const e of entries) if (e.isIntersecting) {
e.target.classList.add("is-visible");
io.unobserve(e.target);
}
}, { threshold: 0.3 });
document.querySelectorAll("[data-reveal]").forEach((el) => io.observe(el));
</script>
MOTION MUST BE MOTIVATED. Before adding any animation ask: what does it
communicate? Valid: hierarchy, storytelling sequence, feedback, state
transition. Invalid: "it looked cool." If you can't articulate the reason in
one sentence, drop it. Not every card needs an infinite loop — informational
sections stay still.
"Motion claimed, motion shown." If MOTION_INTENSITY > 4, the page
actually moves: hero entry, key-section reveals, CTA hover physics at
minimum. Can't ship working motion in scope? Drop the dial to 3 and ship a
clean static page. Never half-build motion that breaks.
MARQUEE MAX-ONE-PER-PAGE. Two marquees on one page is lazy filler.
window.addEventListener("scroll", …) is banned. Use scroll-driven CSS,
IntersectionObserver, or nothing. Same ban for rAF loops recomputing layout
from scrollY.
Animate ONLY transform/translate/scale and opacity. Never top,
left, width, height. will-change sparingly.
Reduced motion is non-negotiable. Anything above MOTION_INTENSITY 3
honors prefers-reduced-motion: reduce: infinite loops, parallax, and
reveals collapse to static/instant.
Grain/noise overlays live exclusively on a fixed pointer-events: none
element — never on scrolling containers (continuous repaints destroy mobile
FPS).
z-index restraint. A documented scale for systemic layers (sticky
header, modal, scrim, grain) — no arbitrary z-index: 50 sprinkled around.
8. Dark mode & theme lock
Hwaro's token system (Section 2) makes both schemes nearly free — the
discipline is what remains:
PAGE THEME LOCK. The page has ONE theme: light, dark, or auto
(color-scheme: light dark). Sections never invert mid-scroll — no
warm-paper section sandwiched between dark sections. Section-level tints
within the same family are fine (--bg next to --bg-subtle); a full flip
is broken. Exception: one deliberate full theme-switch device per page,
only when the brief explicitly calls for it.
Hierarchy parity. If the CTA pops in light, it pops in dark. The brand
accent stays recognisable in both — don't desaturate the brand into dark
mode.
No pure #000000 / #ffffff — the ember ink ramp and surfaces model
this; keep custom pairs off-black/off-white too.
Test both modes before finishing. Screenshot light AND dark (Section 11).
Never ship a scheme you haven't seen.
9. AI Tells (forbidden patterns)
Avoid these signatures unless the brief explicitly asks for them. These came
out of real production tests of LLM-generated pages — they are what the model
does when it tries to "look designed."
9.A Visual & CSS
NO neon / outer glows by default — inner borders or subtle tinted shadows.
NO pure black #000000 — off-black / charcoal (the token ramp).
NO oversaturated accents; desaturate to sit with the neutrals.
NO gradient text on large headers as a default move.
NO custom mouse cursors — outdated, a11y-hostile, perf-hostile.
NO crosshair / hairline grid lines as pure decoration — hairlines organize
real content or don't exist.
NO version labels in the hero (V0.6, BETA, EARLY ACCESS) unless the
brief is literally a launch/preview announcement.
NO "Brand · No. 01"-style micro-meta sub-eyebrows.
NO decoration text strip at hero bottom (DESIGN · BUILD · SHIP,
ESTD. 2018 · LISBON) — unless the strip carries real navigable links or
real status.
9.D Section numbering & micro-labels
NO section-number eyebrows (00 / INDEX, 001 · Capabilities,
06 · how it works). Eyebrows name topics in plain language — and are
rationed anyway (Section 4.6).
NO 01 / 4 pagination labels on images or bento tiles.
NO generic step labels ("Step 1 / Step 2 / Step 3", "Phase 01…"). The verb is
the label: "Install", "Configure", "Ship".
NO micro-meta sentences under eyebrows ("Each of these is a feature we ship
today, not a roadmap promise."). Eyebrow + headline + body is enough.
9.E Separators, dots, and flourishes
The middle-dot (·) is rationed: max 1 per line in metadata strips,
never the default separator for everything.
ZERO decorative status dots by default. A colored dot before nav items /
list rows / badges is a tell. Only for real semantic state (live status), at
most one per section.
NO <br>-broken-and-italicized headlines as a default move — headlines read
naturally first.
NO vertical rotated text — agency cliché; only when the brief is explicitly
experimental AND it serves the composition.
9.F Pills, captions, stamps
NO pills/labels/tags overlaid on images (PLATE · BRAND, Field notes — journal). Let the image speak, or caption below it.
NO photo-credit captions as decoration (Field study no. 12 · Ines Caetano)
— credit only real photographers on real photos; otherwise a one-line
functional caption or nothing.
NO version footers on marketing pages (v1.4.2, Build 0048,
last sync 4s ago). CLI fixtures, not landing content.
NO fake live-stock counters ("Reservation 412 of 800") without real data.
9.G Marketing-copy tells
NO "Quietly in use at / Quietly trusted by" — say "Trusted by", "Used at",
or let the logos speak.
NO poetic section labels ("From the field", "Currently on the bench", "Loose
plates") — plain functional labels ("Testimonials", "Latest writing") or
none.
NO mock-humble industry references ("We respect the French ones").
NO locale / city / time / weather strips ("LIS 14:23 · 18°C") for 99% of
briefs — allowed only for genuinely place-focused or globally-distributed
briefs. A contact address in the footer is fine; an atmospheric locale strip
is not.
NO scroll cues ("Scroll", "↓ scroll to explore", animated mouse icons). The
user knows what scrolling is.
9.H Lists, dividers, scoring
NO top and bottom borders on every row of a long list/spec table — one
divider direction, used sparsely; better, use a different component
(Section 4.7).
NO scoring/progress bars with filled background tracks as comparison visuals
on marketing pages — a number + small icon, or a tiny trackless bar.
9.I EM-DASH BAN (the single most-violated tell)
The em-dash (—) is completely banned in visible page copy. No "sparingly"
allowance — historically "use sparingly" gets ignored, so the rule is binary:
zero.
Banned in headlines, eyebrows, labels, pills, buttons, captions, nav items,
body copy, quotes, attribution, and alt text.
Restructure instead: two sentences, a comma, parentheses, or a colon.
The en-dash (–) as a separator is banned too. Date ranges (2018-2026) and
number ranges (€40-80k) use a plain hyphen.
Permitted dashes in page copy: the hyphen (-) and the math minus (-5°C).
This governs the site's visible copy (templates and authored demo content
you write). If a single — or separator-– is visible on the page, the
output fails pre-flight and gets rewritten.
10. Redesign protocol
This skill handles greenfield builds AND redesigns. Misclassifying the mode is
the biggest source of bad redesign output.
10.A Detect the mode (first action)
Greenfield — no existing site, or full overhaul approved.
Redesign — preserve — modernize without breaking the brand. Audit first,
extract brand tokens, evolve gradually.
Redesign — overhaul — new visual language over existing content. Treat as
greenfield for visuals; preserve content and IA.
If ambiguous, ask once: "Should this redesign preserve the existing brand,
or are we starting visually from scratch?"
10.B Audit before touching
Document the current state first:
Brand tokens — the existing :root values (or, on a non-token site, the
de-facto palette/type/radii you extract from the CSS).
Information architecture — section tree, nav, key paths.
Content blocks — what exists, what's doing work, what's filler.
Color recalibration — desaturate, unify neutrals, keep the brand accent.
Motion layer — dial-appropriate micro-interactions on existing
components.
Hero & key-section recomposition — using the Section 6 vocabulary.
Full block replacement — only when a block is unsalvageable.
Targeted evolution (levers 1–4) captures ~70% of the value at ~40% of the risk
when IA and content are sound. Go full redesign only for structural visual
debt; go greenfield only when the brand itself is changing.
11. Workflow
Design Read + dials (Sections 0–1). Infer when you can; short question
round only when the read genuinely diverges; play back the brief for
substantial engagements.
Propose direction(s). Offer 1–3 distinct concepts (not three shades of
one idea), each as a short pitch: the organizing idea, the palette (token
values), the type pairing, the layout move, the motion stance. Make each
concrete with a token-value block or ASCII sketch — AskUserQuestion
renders these as side-by-side text previews. Save real visual comparison
for the browser (step 4). For a confident single read, one declared
direction is fine — state it and build.
Implement, token-first. Define/extend the :root token layer, then
style components against tokens, then assemble layouts in templates. Keep
markup semantic and accessible from the start. Internal URLs carry
{{ base_url }} (Section 2).
Iterate live — and actually look at it. Run hwaro serve and refine
with the browser open. If you can't see a browser, build and screenshot
with headless Chrome (--screenshot) and review the image — never judge
a visual design from HTML/CSS source alone. Screenshot at a mobile width
(~390px) AND a desktop width (~1440px), in BOTH color schemes. Hand the
user the hwaro serve URL early; design is a conversation.
Pre-flight, then ship. Run the full Section 12 check, fix every failing
box, then present the result and explicitly invite feedback ("warmer /
denser / quieter?"). Don't declare it done — let the user.
Two paths
Retheme a scaffold (fast path).hwaro init my-site --scaffold blog
(also simple / docs / book; bare ships no CSS at all) gives you a
complete, accessible, token-driven baseline that already adapts to light and
dark. The *-dark variants are the same sheet with color-scheme: dark
pinned — presets, not separate designs. Often the highest-leverage move is
rewriting the token pairs and a few component rules to the user's brief —
minutes to a distinctly different site. For a custom dark (a warm dark, not
stock ember-dark), override the dark side of the light-dark() pairs —
the syntax theme follows via --code-*. Remember Section 0.D: a retheme
still needs a Design Read; shipping stock ember unchanged is not a design.
Build bespoke. When the brief needs a layout the scaffolds don't have,
author templates from scratch (or start from --scaffold bare/simple) and
bring your own token system. More work, full control.
12. Final Pre-Flight Check
Run this before presenting the result. This is not optional. If any box
fails, the output is not done — fix it first.
Read & direction
Design Read declared (one line) with reasoned dial values, not silent
defaults?
Redesign mode detected and audit performed (if applicable)?
Result matches the confirmed brief, or the deviation was discussed?
One coherent point of view visible throughout — and NOT stock ember
passed off as a custom design?
Tokens & color
Everything reads from tokens; no orphan hardcoded colors (scaffold
hygiene spec stays green when editing scaffold source)?
Every custom color pair defines BOTH light-dark() sides?
One accent, used identically across all pages/sections (Color
Consistency Lock)?
Neutrals are off-black/off-white, never pure #000/#fff?
Premium-consumer brief? Palette is NOT the beige+brass+espresso family
(i.e., not just re-shipped ember warmth)?
Page Theme Lock: one theme, no section inverts mid-page?
--code-* syntax slots tied to the palette in both schemes?
Typography
Type scale + line-height + measure deliberate; pairing intentional; CJK
fallbacks preserved where the audience needs them?
Serif discipline: landing/portfolio brief didn't default to serif (and
never Fraunces / Instrument Serif without explicit brand justification)?
Emphasis inside headlines is same-family italic/bold, not mixed-family?
Italic display words with descenders have line-height ≥ 1.1 + padding
reserve?
Fonts self-hosted, subset, font-display: swap?
Layout discipline
Hero fits the viewport: headline ≤ 2 lines, subtext ≤ 20 words, CTA
visible without scroll, 100dvh not 100vh?
Hero stack ≤ 4 text elements; no tagline/trust-strip/pricing teaser in
the hero; logo wall UNDER the hero?
