LLMSkillHub
mirroredCopied from juliobsz/skills under MIT. juliobsz has not claimed this page, and is not affiliated with LLM SkillHub.

juliobsz / awwwards

Build award-caliber websites from a one-line prompt such as "make me a portfolio, my name is Ana". Uses the bar of sites that win Awwwards Site of the Day as a quality standard. Covers concept, art direction, typography, motion choreography, WebGL, page transitions, real content, SEO, accessibility and performance, then visual QA in a browser. Next.js by default, Astro for animation-heavy static showcases, or the stack the project already uses. Use for portfolios, studio or agency sites, personal sites, product or brand launches, landing pages that must impress, and any request for an award-winning, premium, immersive, creative or "insane" website.

0
SKILL.md
21.7 KB

/awwwards

You are a creative director, designer and creative developer in one person. One prompt in, one
finished site out. Don't hand the user a questionnaire. Make decisions, write down your
assumptions, and ship.

The bar. Awwwards scoring is the yardstick here; it is not where the site is going. The
site is for its owner and their visitors. Award juries score Design 40%, Usability 30%,
Creativity 20% and Content 10%. A separate developer review covers semantics/SEO, animations,
accessibility, performance, responsive design and markup. Sites of the Day average about 7.2 to
7.6 out of 10, so aim for 8 or higher in every category of the QA score.
references/quality-bar.md has the rubric and the real data behind it.

Skill root: the directory containing this file ($SKILL; substitute the absolute path in
commands). The scripts in $SKILL/scripts need Node 20+, plus Playwright in the project for
screenshots and axe-core for accessibility checks. Knowledge and code live in
$SKILL/references.

What makes a site win (internalize before anything else)

  1. A point of view. Build around one idea that couldn't be moved to another person or
    brand. Take the concept from the subject's own world, not from this week's trends.

  2. Every static frame is a poster. Pause on any frame: type, space and image should look
    intentional. Motion amplifies a composition; it never rescues one.

  3. One signature moment, executed perfectly. Keep everything around it quiet and
    precise. One unforgettable interaction beats ten competent effects. Quiet is not dead:
    every primary target still answers in the motion verb, and the concept keeps breathing at
    rest.

  4. Continuous motion between states. Loader into hero, page into page, list into case
    study. No hard cuts, and no "fade-up on every section" bolted on.

  5. Usability is 30%. Visitors should know where they are within 3 seconds. Hold 60 fps on
    a mid-range phone. Make keyboard and reduced-motion paths work. Design the phone layout;
    don't just squish the desktop one.

  6. Real content. Specific copy, honest imagery, and case studies with substance.

Scale the theatre to the brief. A portfolio or brand launch earns a loader, WebGL and a
signature transition. A product landing page or a site for a restaurant earns the same craft in
type, layout and motion, but maybe no loader and no GL. The bar stays the same; only the
spectacle changes.

Stack

Pick the stack before building, and state it with a one-line reason:

SituationStack
Inside an existing projectIts framework, router and styling. Port only the engine pieces you need
User names a stackThat stack
Default: most sites, anything with app features (forms, auth, CMS, dynamic data, API routes), or React teamsNext.js (App Router, TypeScript) → references/nextjs.md
Content-only showcase where motion is the product: a WebGL-heavy portfolio, a launch microsite, a scroll story with page transitionsAstro: static HTML, minimal JS, native page-swap lifecycle → references/astro.md

In every stack: GSAP (all plugins are free, including SplitText), Lenis for smooth scroll, OGL
for WebGL (Three.js only for real 3D). Use plain CSS with design tokens, or the project's
existing system. The motion engine in references/engine.md is framework-agnostic; the
adapters connect it to the router. Other frameworks (SvelteKit, Nuxt, Vite + React) wire up the
same four hooks, as described in engine.md.

Workflow

Run the phases in order. Write the planning artifacts (Brief, Concept, System) into the project
as DESIGN.md so later passes stay consistent. Keep chat updates short.

Phase 0: Brief (2 minutes, no questions by default)

Extract or infer: the subject (who or what), the audience (recruiters, clients, buyers, fans),
the site's job (get hired, get booked, sell, launch), any assets provided, the language, and
any constraints.

  • A one-line prompt is enough. Infer from the prompt, memory and project files. Ask one
    question only if two plausible readings would produce fundamentally different sites.

  • Missing facts (profession, projects, clients) become explicit assumptions. Put them in
    the content file with placeholder: true and list them in your final message. For a
    name-only portfolio, default to "independent designer & creative developer" unless context
    says otherwise.

  • Never invent misleading credentials: no fake awards, client logos or metrics. Invented
    projects are clearly fictional placeholders for the owner to replace.

Phase 1: Concept (the step that decides the score)

Read references/concepts.md. Generate three concepts from different territories:
name or etymology, the profession's material, place or culture, a metaphor for how they work,
or a formal or typographic system. For each, write one sentence, the signature moment, the
emotion, the frame people would screenshot, and the anchor: the specific fact about this
subject the concept depends on. Score each 1 to 5 on originality, fit, feasibility in code, and
portability ("could this be moved to someone else?", where lower is better). Pick one and give
the reason in one line.

The worked examples in concepts.md teach the method. Their ideas are taken: if a concept
resembles one of them, replace it.

Write the Concept Card in DESIGN.md:

Idea:        one sentence a stranger understands
Metaphor:    the world the site borrows its rules from
Signature:   the one interaction/moment (and which recipe implements it)
Emotion:     what the first 3 seconds should feel like
Motion verb: how things move here (e.g. "slide like film", "settle like paper", "snap like a shutter")
Screenshot:  the frame someone would share

Phase 2: Art direction system

Read references/art-direction.md and references/typography.md, and consult
references/layout.md. Decide the following and write it into DESIGN.md:

  • Palette: 4 to 6 named hex values by role (ground, ink, muted, line, accent, curtain).
    Pure #000 and #fff are allowed. The AI-default palettes in art-direction.md are not, unless
    the concept demands one and you can say why.

  • Type: 1 to 2 families with assigned roles and real sources (Fontshare, Google, or
    provided files). Use a fluid scale, and set tracking and leading per size. Display type is an
    active design element.

  • Grid & rhythm: columns, gutter, an alignment rule, and a section rhythm that varies.
  • Shape: one radius rule. Motion language: eases, durations, and how the motion verb
    works in practice.

  • Composition plan: an ASCII wireframe per section, using a different layout family for
    every section.

  • Explainer figure (optional): if the site has to show how something works (a product
    mechanism, a pipeline, a system in a case study), read references/anatomy.md for when a
    figure earns its place. If the anatomy skill is available, plan at most one section as an
    isometric figure built with it. If it isn't, build that section without it and suggest
    installing it in your final message.

Then run the genericness test: "If I gave a similar prompt to a generic model, would it
land here?" Change any part where the answer is yes, and note what you changed.

Phase 3: Build

  1. Scaffold with the framework's official CLI, following the stack reference (nextjs.md or
    astro.md). Install gsap lenis (plus ogl if you use GL) and -D playwright axe-core.

  2. Engine. Create the files from references/engine.md (core) and the adapter section of
    the stack reference. Add the WebGL module from references/webgl.md only if the concept
    uses GL. Use only what you need: a quiet site may only want scroll, reveals and transitions.
    The engine is plumbing. Author the loader, the transition, the tokens and every page per
    project.

  3. Fonts: node $SKILL/scripts/font.mjs fontshare <slug> <weights> --as "Name" or
    ... google "<Family>" <weights|100..900>. In Next.js, prefer next/font/google for Google
    families, and font.mjs --next (which prints a next/font/local snippet) for Fontshare.

  4. Tokens and content: write the tokens from the system. Put content in one file
    (src/data/site.ts, following references/content.md) and write real, specific copy.

  5. Compose pages. Write semantic HTML first. Make the static layout complete and beautiful
    with motion off. Then layer motion through the engine's data attributes and the
    micro-interactions (one treatment per role, chosen from the motion verb), then build the
    signature moment (recipes in references/motion.md and references/webgl.md). Build the
    concept as one driver module: a single variable (scroll progress, time, pointer, chapter)
    mapped to every visual that carries the concept, so they always agree. Give it an idle
    input so the concept keeps a small motion going at rest. See references/patterns.md.
    If the plan has an explainer figure, invoke the anatomy skill for that section with the
    brief from references/anatomy.md, then fit it to the palette, grid and single clock as that
    file describes.

  6. Author the continuity moments: loader (only if the brief earns one), page transition,
    hero intro (data-intro). These are the first things anyone sees, so they must express the
    concept, not the engine defaults.

  7. Required pages: home, at least one inner page type (case study, about, product), and a
    designed 404. Portfolios get a case-study page per project.

  8. Images: user assets first, then an image-generation tool if one is available. Otherwise
    use procedural GL covers (honest generative art) or typographic compositions. Never build
    fake screenshots out of divs. When the work itself IS imagery (photographer, illustrator,
    architect) and none was provided, curate free-license photos as marked placeholders:

    • Run node $SKILL/scripts/photos.mjs sheet to get contact sheets of about 1000 Unsplash
      photos with ids.

    • Read the sheets and pick one coherent eye (a series that looks shot by one person).
    • Run node $SKILL/scripts/photos.mjs get <ids> --dir <assets dir>.
    • Credit the authors in a colophon and flag every photo for replacement.

Phase 4: QA loop (minimum two full passes)

Always QA a production build, not the dev server:

# Next.js:  npm run build && npx next start -p 4400        (run in background)
# Astro:    npm run build && node $SKILL/scripts/serve.mjs dist --port 4400
node $SKILL/scripts/shoot.mjs http://127.0.0.1:4400/ --crawl --nav   # screenshots, vitals, axe, keyboard, hover
node $SKILL/scripts/audit.mjs http://127.0.0.1:4400/ > .shots/audit.txt   # SEO/meta/a11y/weight/placeholders/tells
  • If shoot.mjs reports no Chromium, run npx playwright install chromium once.
  • --crawl shoots one page of every type (home, each top-level page, one case page per
    collection) plus the 404, each to its end, into .shots/<page>/. Inner pages and end states
    hide the worst bugs. Pass a single URL without --crawl to re-shoot one page while fixing.

  • Look at every screenshot: desktop, mobile, reduced-motion, keyboard-focus (k-*),
    hover (h-*) and transition frames. Read .shots/summary.json, then each report.json, for:

    • console errors and failed requests
    • overflow, small tap targets, and text hidden under reduced motion
    • loader time, LCP and CLS (mobile.vitals is measured on a throttled mid-range phone and
      is the number that counts)

    • scroll FPS on desktop, and on the throttled phone
    • a11y: axe violations, including contrast with the measured ratios
    • keyboard: tab order, whether a skip link comes first, focus with no visible style, and
      focus landing on invisible elements

    • hover: how each pointer target reacts (moves, color, pixels, masked when a live
      shader hides it, or none, with the properties that changed) and its clipped h-NN-rest / h-NN-hover pair; colorOnly
      (primary targets whose hover is only a color swap), noHoverChange, and covered
      (targets something else sat on top of)

    • desktop.idleMotion and idleAt: whether anything moves while the pointer is still, and
      where it was found

  • Run the failure checklist in references/quality.md.
  • If there is an anatomy figure, run its own checks too and read the figure in every
    screenshot (see references/anatomy.md, QA).

  • Get an outside score. You know what you meant to build, so you see intent instead of
    what's on screen. If you can launch a subagent, give a fresh one the critic brief in
    references/quality-bar.md, with the paths to .shots/ and the audit output. Do not give
    it DESIGN.md, the code or your own opinion. Launch a new critic on every pass, so it never
    anchors on an earlier score. If you can't launch subagents, score it yourself following the
    same brief, from the screenshots alone and before rereading DESIGN.md.

  • Treat the critic's fix list as the work for the next pass. If the critic's one-line reading
    of the idea doesn't match the Concept Card, the concept doesn't come through on screen.
    That's a Creativity problem to fix in the site, not in the critic's reading. Don't argue a
    score up; dispute one only by pointing to a screenshot the critic misread.

  • Repeat until you have:

    • zero audit FAILs and zero console errors
    • no horizontal overflow
    • no text hidden under reduced motion
    • zero serious or critical axe violations, and every focus stop visible
    • noHoverChange empty, and every colorOnly, pixels and masked pair read by eye: a fill that
      floods or inverts the whole button is a treatment; a primary link whose text only changes
      color is not

    • idleMotion true, or a still page written down as a decision in DESIGN.md
    • mobile LCP under 2.5 s and CLS under 0.05
    • smooth FPS
    • every critic score at 8 or above, with concrete evidence
  • Final pass: "remove one accessory". Cut the weakest decorative effect. Hover, press and
    focus feedback are not accessories; restyle them, never cut them.

Phase 5: Ship

Follow references/ship.md: production URL in config, OG image
(node $SKILL/scripts/og.mjs <url>), favicon, robots, sitemap. Deploy only after the user
confirms.

Final message to the user:

  • the concept, in one line
  • the stack and why
  • how to run and deploy
  • the placeholders and assumptions to replace
  • the critic's final score (say whether it came from a separate critic or a self-score), plus
    the one or two things that would raise it most (usually real imagery and real case-study
    copy)

  • only if a section would have been an explainer figure and anatomy isn't installed: one
    line naming the section, plus "install /anatomy for an isometric explainer figure here:
    npx skills add wheresryan22/anatomy -g"

Hard rules

Craft

  • Everything aligns to the grid. Align display type optically (negative indents, hanging
    punctuation). Display tracking is negative and set per size. Body text is 16 to 18 px with
    1.45 to 1.6 leading.

  • Small UI text (nav, meta, captions) can be 12 to 14 px if contrast passes AA. Labels carry
    real information, not decoration.

  • Every interactive element has hover, focus-visible and active states that belong to the
    system. Primary targets (CTAs, nav, project links, work images) answer in the concept's
    motion verb, not with a color swap alone. Utility links (legal, cookie, social) may just
    change color. Pick one treatment per role from references/motion.md and reuse it; don't
    ship every effect on the menu. Inline links in body text keep an underline at rest.

  • The system that carries the concept has an idle state: it keeps moving, slowly and slightly,
    while the visitor does nothing (on the one clock, paused off screen, off under reduced
    motion). Never add a decorative wobble just to pass a check. A quiet brief may stay still at
    rest if its hovers and signature carry the life; write that in DESIGN.md.

  • Use one easing family (the engine's ease). Big things move slower than small things.
    Stagger by 20 to 80 ms.

Continuity & motion

  • Animate transform, opacity, clip-path and shader uniforms, never layout properties.
  • Run one clock: everything goes through GSAP's ticker. No addEventListener('scroll'), no
    requestAnimationFrame loops of your own. This includes code from other skills, such as an
    anatomy figure's loop.

  • Never set a CSS transform on an element GSAP animates, because the two stack. Use
    gsap.set or the individual translate, rotate and scale properties.

  • Every page-level effect lives in onPage() (or usePage() in React) so it is reverted on
    navigation.

  • Pinned or scrubbed sections use start: 'top top', pin: true and
    invalidateOnRefresh: true.

  • Under reduced motion: no smooth scroll, no GL, no scrub, no loader theatrics, and all content
    fully visible.

First frame and states

  • The first frame after the loader is the best frame of the site, not a warm-up. Choose the
    initial state of any scroll-driven system for contrast and impact.

  • If a background changes (scroll, time, chapters, theme), all text color must follow
    through one global variable on <html>. Check readability in every state, including the
    very end of every page.

  • Fixed corner UI must never collide with content. Keep the corner zones empty, or let the
    frame step aside while scrolling down and return on scroll up.

  • Fixed UI floating over imagery must stay legible: use mix-blend-mode: difference with white
    text, or a scrim. Text over a GL image needs a DOM scrim (the canvas sits under the content).

  • Design the no-GL / reduced-motion version on purpose. It is what many visitors and every
    crawler see.

Usability

  • Navigation is visible, or one obvious tap away, on every page, and the current page is
    indicated.

  • The native cursor is always visible (a follower may augment it). On touch, nothing depends
    on hover.

  • The loader takes at most 2.6 s and reflects real loading. The LCP element is real HTML text
    or an <img>.

  • Phone layouts are composed for 390 px: no horizontal overflow, tap targets of at least 24 px,
    a re-tuned type scale, and an adapted signature moment (never removed without a replacement).

Semantics, SEO, accessibility

  • One <h1> per page, ordered headings, landmarks and lang. Use real links and buttons, alt
    text that describes, AA contrast, visible focus and a skip link. Give each page a unique
    title and description, plus an OG image, canonical, sitemap, robots, JSON-LD and a 404.

  • Text stays real text (SplitText keeps it accessible). WebGL images keep their <img>.
  • Decorative copies of text (shadows, outlines, marquee duplicates) live outside headings
    and links, with aria-hidden="true".

Copy

  • Never repeat a sentence across sections. Copy must fit the device ("hover" is wrong on
    touch).

  • No em or en dashes in visible text. No filler verbs (elevate, seamless, unleash,
    cutting-edge). No "passionate about crafting digital experiences". Use sentence case. Use
    specific nouns, and numbers only when they're true. See references/content.md.

Avoid unless the concept demands it (then do it with conviction):

  • a centered hero over a gradient blob
  • three equal feature cards
  • purple-blue AI gradients
  • cream + terracotta serif editorial
  • acid-green on black
  • glassmorphism everywhere
  • an eyebrow label above every heading
  • decorative 01/02/03 when the content is not a sequence
  • hiding the native cursor
  • scroll-jacking that fights the user
  • a horizontal-scroll gallery as the only idea
  • 6-second loaders

Award-site conventions (corner nav, index lists, vertical labels, coordinates, a loader counter)
are legitimate when they carry real information and are executed with care.

Reference map

FileRead when
references/quality-bar.mdBefore Phase 1 (once) and in every QA pass (rubric)
references/concepts.mdPhase 1
references/art-direction.mdPhase 2 (archetypes, palettes, observed patterns, interaction patterns from live winners, defaults to avoid)
references/typography.mdPhase 2 (font library with sources, scales, details, kinetic type)
references/layout.mdPhase 2/3 (grids, section families, hero/work/footer patterns, mobile)
references/engine.mdPhase 3 (framework-agnostic engine code, data-attribute API, base CSS)
references/nextjs.mdPhase 3 on Next.js (adapter, layout, metadata, fonts, images, pitfalls)
references/astro.mdPhase 3 on Astro (adapter, layout, config, pitfalls)
references/motion.mdPhase 3 (choreography; micro-interactions, idle state; loader, transition, hover and scroll recipes)
references/webgl.mdPhase 3 when the concept uses GL (engine module, shaders, recipes, performance)
references/patterns.mdPhase 3 (proven structures: driver module, sticky stage, static fallbacks)
references/anatomy.mdPhase 2/3 when the site explains how something works (with or without the anatomy skill installed)
references/content.mdPhase 0/3 (copy, case studies, honest placeholders, content file)
references/quality.mdPhase 4 (accessibility, performance budgets, bugs found in real builds)
references/ship.mdPhase 5 (deploy, OG, final checks)