award-grade

An award-grade AI design agency, packaged as composable Claude skills — with an adversarial judge that can refuse to ship.

Eight skills that give any Claude agent the working methods of a world-class digital studio — signature interactions, motion choreography, WebGL effects, design-DNA extraction, programmatic video, narrative UX — all governed by one shared canon and gated by an adversarial award judge. Built to be used, forked, and learned from.

The thesis: "wow" is not an effect you add. It's one loud idea, ten quiet perfect decisions, a 60fps floor, and the discipline to cut everything else. This suite encodes that discipline so an AI agent can practice it.

The agency

The canon (what makes this a studio, not a folder of tips)

Every skill obeys four shared laws in references/:

• motion-canon.md — timing/easing tokens, choreography rules, and the slop bans (transition: all, fade-in-up-everything, bounce-on-product-UI…)
• awwwards-scoring.md — the jury rubric (Design 40 · Usability 30 · Creativity 20 · Content 10), the memory gate, the generic-AI tells
• performance-budgets.md — 60fps floors, LCP/CLS/INP budgets, WebGL payload caps, the test ritual
• accessibility-motion.md — reduced-motion as art direction, vestibular safety, keyboard integrity under wow

The judge enforces the canon. Canon violations cap scores. That's the whole trick: taste, made mechanical enough to gate.

Eyes, memory, honesty (what makes it agent-native)

Doctrine alone can't stop an agent from guessing — so the suite ships the machinery:

• Eyes — scripts/judge-evidence.mjs: one command collects the judge's evidence pack (screenshots at 1440/768/375 + per-section, a reduced-motion pass, a 4×-throttled trace with long-task extraction, console/network logs, Lighthouse/axe when installed) and auto-fills evidence/LEDGER.md. Collectors degrade honestly: a missing tool writes a MISSING row with the exact human command, never a silent skip.
• Memory — references/taste-memory.md: every verdict appends a machine-readable block to your project's design/TASTE.md; the director reads it at boot. Protect-lists persist, settled arguments stay settled, re-judges diff against the ledger instead of vibes. The agency learns your taste.
• Honesty — references/honest-evidence.md: every verification ritual is labeled [MEASURED] / [EMULATED] / [HUMAN]. An agent can't watch a render at 1× or be five test users — so those lines become explicit human handoffs, not fabrications. Unlabeled = HUMAN-pending, never a fact.
• Canon as data — references/canon.json: the ban list and tokens in machine-readable form; the prose canon and the enforcer both derive from it, so they cannot drift.
• Self-enforcement: scripts/suite-lint.mjs lints the suite itself (frontmatter, dead links, version agreement, duplicate triggers) and scripts/test-canon-audit.mjs proves every audit rule against the paired fixtures/ (violation + fix per rule — currently 70 checks). The repo obeys its own laws in CI.

The craft library (deep technique references)

Beyond the four laws, references/ carries battle-tested technique docs — distilled from real award-site calibration work (76+ phases of recreating production particle heroes, verified with pixel probes):

• threejs-webgl-craft.md — the 12 immutable shader rules (the r152 color-space trap, normalization-constant hunts, render() sanity greps, rAF-not-tween uniforms, the three-part cursor sync…), pixel-probe forensics, library decision matrix, hero-source technique tiers
• gpgpu-particles-and-morph.md — FBO ping-pong particle systems, curl-noise flow fields, mesh morphing as a spring force into the field (never a lerp), and when NOT to GPGPU
• ascii-dot-matrix-art.md — the text-as-image aesthetic family: four-tier technique ladder, the "resolve" as signature move, dot-density data honesty, the full pairing checklist
• gsap-recipes.md — timeline/label architecture, ScrollTrigger scrub-vs-toggle discipline, the React Strict Mode traps, SplitText bounds, Flip, the one-ticker rule
• pattern-gallery.md — 20 NAMED award-site pattern classes (the resolve, ledger-settle, cursor lens, the confident static…), each with real-world sightings, a spec skeleton, and a mandatory divergence line — grammar, not a menu (two per project, max)
• sound-design.md — the earn-it test, user-initiated/mute-first law, and why sound is a reference here, not a ninth skill

Install

As a Claude Code plugin (the whole agency at once):

claude plugin marketplace add <repo-url>      # the repo ships .claude-plugin/marketplace.json
claude plugin install award-grade

Manually (pick your specialists): copy any skills/<name>/ directory into ~/.claude/skills/ — each skill is self-contained plus the shared references/ folder (copy it alongside, or keep the relative links intact by copying the repo whole).

Use

Start every project through the director:

"Use wow-director: a landing page for [product]. It should feel like [one feeling]. Budget: [time]."

The director interrogates the brief, proposes three concept territories, forces the signature moment choice, staffs the specialists, and runs award-judge before anything ships. Or go direct when you know what you need: "use motion-choreography to build a motion token sheet for a calm fintech brand."

Prompt it well

The skills fire on trigger phrases, but output quality is set by prompt quality. PROMPTS.md is the cookbook: the four-ingredient brief anatomy (product / ONE feeling / the 24-hour memory / scope), per-skill patterns, combo pipelines (the calibration loop, the restraint pass, judge-gets-the-last-word), and the anti-pattern table — every slop prompt paired with its rewrite.

Scaffold your project for it

references/project-scaffolding.md — the codebase changes that make the craft survive: token files as law, the /lab isolation route for signature-moment prototypes, design/ artifacts versioned with code, the perf log, and scripts/canon-audit.mjs — a zero-dependency guard driven by canon.json that FAILS builds on the mechanical slop tells — transition: all, entrance theater (3+ entrance animations; a single signature entrance only warns), bounce/overshoot easing (keywords AND numeric cubic-bezier y-values outside [0,1]), inline GLSL, half-pixel type — and WARNS on raw durations outside token files, standing will-change, and animating without ever querying prefers-reduced-motion (--strict promotes warnings):

node scripts/canon-audit.mjs --src src app     # wire into prebuild/CI; --strict promotes warnings

The showcase — judged by its own judge

docs/ is Meridian, the first-mission brief built for real by the suite, then scored by an independent award-judge run using judge-evidence.mjs — receipts committed in showcase/ (evidence packs, BRIEF.md, the full verdict ledger in TASTE.md).

The judge still refuses to say SHIP: real-device pass, screen-reader tab-through, and the 5-person test are [HUMAN]-pending rows it will not fabricate. That refusal is the product working.

Learn

Read examples/first-mission.md — a complete fictional engagement (a coffee roastery, one week, subscription conversions) walked through every skill: the brief interrogation, the three territories, the moment that won, the WebGL idea that was correctly rejected, the first judgment that lands as BLOCKED on a real contrast-gate failure, and the fixes that earn a scored 7.6 — Site-of-the-Day band. The example teaches the restraint as much as the craft.

Design principles behind the suite

1. One signature moment. Juries remember one thing. The pipeline forces you to choose it early and spend on it disproportionately.
2. Restraint is a skill, not an absence. Every skill documents when NOT to use itself. The strongest deliverable in the example is a rejection.
3. Evidence over vibes. No score without a screenshot. No "it works" without recorded numbers. No copying an AI's guess of how an effect works.
4. The reduced-motion experience is a first-class art direction, designed in parallel — not an off switch bolted on at the end.
5. Wow that doesn't hold 60fps on a mid-range phone is lag with ambition.

Extend it

Fork-friendly by design: add a skill directory following the same frontmatter + section pattern, obey the canon, give it a "when NOT to use" section, and wire it into wow-director's staffing table. PRs that add bans to the slop list need a screenshot of the crime.

Credits & lineage

Original content, MIT. The design-DNA three-layer structure builds on a pattern the community has been converging on (notably zanwei/design-dna and Jane-xiaoer's web-clone methodology — both MIT); the evidence-grading discipline (SOURCE/PARTIAL/GUESS) is adapted from the web-clone reverse-engineering school. Scoring weights mirror Awwwards' public jury criteria. Everything here is written fresh for this suite.

License: MIT · Status: v1.4.0 · Built with Claude, judged adversarially by design.