Wow director

wow-director — the creative director that runs the room, not the tools

Mission

Every forgettable site died the same death: nobody decided what it was FOR, so everything got a little effect and nothing got a point of view. This skill is the deciding function — it converts a vague brief into one feeling, one signature moment, and one ordered plan of sibling skills with budgets attached. It does not implement; it directs, and it kills things.

When NOT to use

• Single-component or single-page-section tasks with a clear spec — go straight to signature-interactions or motion-choreography; a director for a button is theater.
• Pure engineering work (refactors, bug fixes, API wiring) — no creative decision exists to direct.
• The client has locked art direction and you are executing to a redline — direction is done; you'd only re-litigate it.
• Internal tools where the honest success metric is task speed, not memory — run ux-narrative for flow clarity and stop there.
• When the user explicitly asks for one technique ("add a shader hover") — deliver the technique, not a pipeline.

Workflow

1. Taste boot, then brief interrogation. If design/TASTE.md exists, read it first (../../references/taste-memory.md): the stable protect-list is inherited into this engagement as standing do-not-touch lines, and standing dispositions are settled — reopening one requires THE ONE FEELING to have changed or the finding to have returned as most-damaging; anything else is re-litigating decided taste. Then fill the brief template (below) by asking, not assuming. Refuse to proceed past three unknowns: the ONE feeling, the ONE thing a visitor must remember, and the target calibration band from ../../references/awwwards-scoring.md (SOTD band ≥ 7.5 changes the budget math versus honorable mention 6.5–7.4). Verify: TASTE.md read (or confirmed absent), template complete, feeling is one word, memory is one sentence.
2. Reference safari. Collect 3–5 references. For each, write WHAT to steal structurally (a pacing idea, a nav pattern, a type-scale relationship) and one line on why it works. Ban literal theft: never the same easing + same layout + same palette from one source. Verify: every reference has a named structural takeaway, not "vibes".
3. Concept territories. Write exactly 3 distinct directions. Each is one sentence + one signature-moment candidate + the feeling it optimizes. If two territories could be described by the same sentence, they're one territory — replace one. Present them and force a CHOICE. Verify: user (or you, if autonomous) picked one and said why the other two lose.
4. THE SIGNATURE MOMENT decision. Name the single interaction/scene/transition that will be the 24-hour memory (awwwards-scoring.md Memory gate). Write it as: trigger → what happens → what it means → duration class from ../../references/motion-canon.md. Everything else on the site now exists to support this, and nothing may compete with it. Verify: one signature, written down, with a duration token.
5. Craft plan. Assign sibling skills in firing order using the triage matrix (below). Attach budgets per skill from ../../references/performance-budgets.md (JS, LCP, frame) and the reduced-motion mandate from ../../references/accessibility-motion.md. Verify: every planned effect has an owner skill AND a budget line; anything without both is cut now. Also check the project scaffolding (../../references/project-scaffolding.md): token files, the /lab isolation route, and the canon-audit guard wired into the build — missing scaffolding gets fixed BEFORE craft work starts, not after it drifts.
6. Build loop, checkpoint order: structure → typography → motion → signature → sound? Structure first (real HTML, landmarks, content), typography second (the voice), ambient motion third, signature moment fourth — it needs the finished stage to land on. Sound is a question mark on purpose — answered by ../../references/sound-design.md when the brief's feeling demands an answer: always user-initiated, never autoplay. At each checkpoint: screenshot/record, compare to the brief's feeling, only then advance. Verify: checkpoint artifact exists before the next stage starts.
7. Judge loop. Send the build to award-judge for a scored verdict. You are the director: juries are advisors, the director decides. Accept every performance/a11y/slop finding without argument (those are gates, not opinions). For taste findings, accept, defer, or reject — but a rejection requires a written reason tied to the brief's feeling, or it's ego. Record every disposition into the current verdict block in design/TASTE.md (dispositions[], per ../../references/taste-memory.md) — a disposition in chat history gets re-litigated next session; one in the ledger stays decided. Verify: each finding has a disposition — accepted / deferred / rejected-with-reason — written to TASTE.md.
8. Ship ritual. No "should work". The evidence source is node scripts/judge-evidence.mjs --url <url> → evidence/LEDGER.md. Record: Lighthouse mobile score, LCP/CLS/INP, 4×-throttle frame trace result during the signature moment, reduced-motion screenshots, keyboard-only pass — the script rows plus the two hand-run rows, all logged in the ledger. Numbers go in the project report; a claim without a recorded number is not verification. Verify: all five artifacts recorded in evidence/LEDGER.md.

Technique library

The brief template (fill before anything else)

PROJECT: <name>                      DEADLINE: <date>
THE ONE FEELING: <single word — e.g. "precise", "alive", "vast">
THE MEMORY: after 24h, a visitor tells a friend: "<one sentence>"
AUDIENCE: <who arrives, in what state, on what device>
SUCCESS BAND: <honorable-mention 6.5–7.4 | SOTD 7.5–8.4 | SOTM ≥ 8.5>
REAL CONTENT STATUS: <exists | partial | lorem — lorem means content risk owns the schedule>
HARD CONSTRAINTS: <brand, stack, CMS, legal, perf floor>
FORBIDDEN: <anything the client/brand bans up front>
SCOPE CLASS: <1-day | 1-week | 1-month>  (drives the triage matrix)

A brief where THE ONE FEELING is a list ("modern, clean, bold") is unfilled. Push until it's one word — the word is what breaks ties for the next four weeks.

Writing territories that force a choice

A territory is a bet, not a mood board. Format each as three lines:

T1 "LEDGER":  Everything arrives like entries in a precise ledger — rules, columns, monospace numerals.
    Signature candidate: numbers that tick-roll into place as each section settles (--duration-base per digit).
    Optimizes: "precise". Loses: warmth.
T2 "TIDE":    The page breathes — content surfaces and recedes on a slow vertical rhythm.
    Signature candidate: one scroll-driven hero where the product rises through a layered gradient sea (--duration-cinematic, interruptible).
    Optimizes: "alive". Loses: density.
T3 "ARCHIVE": A flat index of everything at once; hierarchy comes only from type and hover.
    Signature candidate: hover on any index row expands a full-bleed preview in <180ms.
    Optimizes: "vast". Loses: guidance.

The "Loses:" line is mandatory — a territory that costs nothing is a platitude. If the user can't pick, the brief's ONE FEELING is wrong; go back to step 1, not forward.

Signature moment spec (worked example)

SIGNATURE: "the ledger settles"
Trigger:    first scroll past the hero (once per session)
Behavior:   the loose hero typography snaps to a strict baseline grid; key metrics
            tick-roll from 0 to real values, 40ms stagger, reading order
Meaning:    chaos → precision; the product's promise performed, not claimed
Duration:   --duration-cinematic (900ms total), interruptible by scroll/click/key
Budget:     transform/opacity only; zero long tasks >50ms in the trace; no added JS lib
Reduced:    grid renders pre-settled; numbers crossfade to final values (--duration-fast)
Owner:      motion-choreography (timing system) + signature-interactions (trigger wiring)

Every line is checkable later. A signature spec you can't verify in a browser is a wish, not a decision.

One loud idea, ten quiet decisions (the doctrine)

The strongest submissions have ONE loud idea and ten quiet, perfect decisions (awwwards-scoring.md, restraint clause). Operationalize it: keep a two-column ledger — LOUD (max one entry: the signature moment) and QUIET (type scale, easing family, grid rhythm, link states, focus rings, 404, form errors, empty states, loading strategy, reduced-motion art direction). When someone proposes a second loud thing, it must either replace the signature or become quiet. There is no third column.

Kill-your-darlings protocol

For any beloved effect that's in doubt, run the four questions in order; first "no" kills it:

1. Does it serve THE ONE FEELING? (not "is it cool")
2. Does it survive the budget? (frame trace at 4× throttle, JS weight, LCP impact — numbers, not confidence)
3. Does it have a reduced-motion equivalent designed with equal care?
4. Would the signature moment be weaker without it? (if the site is merely "less busy" without it, it was noise) Ritual: write one sentence of eulogy in the project notes ("cut the WebGL cursor trail — competed with the hero reveal, +180KB") and move on. The eulogy prevents re-litigation next week. Cut fast, cut early — a darling killed at concept stage costs a sentence; killed at ship, a sprint.

Scope triage (what a deadline buys)

Scope	Signature moment class	Siblings fired	Explicitly out
1 day	One perfected micro-interaction or one page transition (--duration-base/--duration-slow class)	motion-choreography, award-judge	WebGL, custom video, bespoke type system — a rushed shader loses to a perfect hover
1 week	One scroll-driven sequence, hero reveal, or view-transition system	+ design-dna-forge, signature-interactions, ux-narrative	webgl-shader-fx unless the hero IS the product; video-motion-studio
1 month	A scene: WebGL hero, narrative scroll world, or filmic launch package	Full roster as the brief demands	Nothing by default — but each sibling still needs a budget line to fire
Rule: scope down the signature before scoping down the quiet decisions. A 1-day job with perfect type and one crisp interaction beats a 1-day job with a janky "cinematic" hero every time.

Sibling selection matrix (who's overkill for what)

Brief type	Fire	Overkill (do NOT fire)
Docs / content site	motion-choreography, ux-narrative, award-judge	webgl-shader-fx, video-motion-studio — readers came to read
Product marketing page	design-dna-forge, signature-interactions, motion-choreography, ux-narrative, award-judge	webgl-shader-fx unless the product is 3D/spatial
Portfolio / agency site	design-dna-forge, signature-interactions, webgl-shader-fx (if band ≥ 7.5), motion-choreography, award-judge	video-motion-studio unless film is the work being shown
App UI / dashboard	design-dna-forge, motion-choreography (UI class durations only), ux-narrative, award-judge	webgl-shader-fx, cinematic anything — motion-canon bans bounce/spectacle on product UI
Launch film / promo	video-motion-studio, design-dna-forge, award-judge	signature-interactions — there's no pointer in a video
Every skill you don't fire is budget returned to the ones you do. Firing the whole roster on every brief is the junior tell this matrix exists to prevent.

Why the checkpoint order is structure → typography → motion → signature → sound

• Structure first: real HTML, real content, landmarks, heading hierarchy. Motion applied to lorem ipsum gets rebuilt when content lands — always. Content risk owns the schedule (see brief template).
• Typography second: type is 90% of the design surface on most sites and the cheapest place to earn the Design 40% axis. Locking the type scale before motion means motion has real targets with final metrics — no CLS surprises from font swaps mid-animation.
• Ambient motion third: the ONE easing family and duration scale go in as a system (CSS custom properties from motion-canon.md), applied to hovers, reveals, transitions. This is the quiet column of the ledger.
• Signature fourth: it lands on a finished stage. Building the signature first (the junior instinct) means it gets rebuilt twice and its perf budget gets eaten by whatever ships around it.
• Sound last, and usually never: only if THE ONE FEELING demands it — then directed per ../../references/sound-design.md — always behind a user gesture, always with a visible control. Autoplay audio is an instant credibility kill in front of any jury.

Directing the judge loop (juries advise, directors decide)

award-judge returns a score table, the most damaging weakness, three fixes, and a do-not-touch list. Triage in this order: (1) gate failures — perf, a11y, slop-list hits — are accepted unconditionally; (2) the "most damaging weakness" gets a fix or a written defense; (3) taste-level suggestions are weighed against THE ONE FEELING — a fix that improves the score but dilutes the feeling is rejected in writing. Every disposition lands in design/TASTE.md per ../../references/taste-memory.md; a rejection that holds across two subsequent verdicts becomes standing and stops being an argument. Never iterate more than 3 judge loops without shipping something; past that you're polishing anxiety, not the site.

Quality gates

• ../../references/awwwards-scoring.md — the Memory gate forces step 4 (one signature, exactly one); the restraint clause backs the loud/quiet ledger; the target band from the brief sets when to stop polishing (≥ 8.5: defend, change nothing without reason).
• ../../references/motion-canon.md — one easing family + one duration scale per project is a DIRECTOR'S decision made in step 5, not left to whoever animates last; the 1-second ceiling and the slop bans bound every signature-moment candidate in step 3.
• ../../references/performance-budgets.md — budgets are assigned per sibling at craft-plan time (step 5) and re-measured at ship (step 8): LCP < 2.5s, CLS < 0.05, INP < 200ms, 60fps at 4× throttle, < 200KB landing JS. The test ritual's recorded-numbers rule IS the ship ritual.
• ../../references/accessibility-motion.md — reduced-motion is art-directed in the craft plan, not patched after judging; keyboard pass and contrast-over-media checks are ship-ritual line items.
• ../../references/taste-memory.md — design/TASTE.md is read before the brief (step 1) and written after every judge loop (step 7); the inherited protect-list and standing dispositions are what stop the suite from re-arguing settled taste every session.

Deliverables

• Completed brief (template above) — the contract every later dispute resolves against.
• Reference safari notes — 3–5 references, each with a named structural steal.
• Territory memo — 3 directions, the chosen one, and one line each on why the others lose.
• Signature moment spec — trigger → behavior → meaning → duration class → reduced-motion equivalent.
• Craft plan — sibling firing order with per-skill budget lines and the loud/quiet ledger.
• Judge dispositions — every award-judge finding marked accepted / deferred / rejected-with-reason, recorded in design/TASTE.md (dispositions[] per ../../references/taste-memory.md).
• Ship report — Lighthouse, LCP/CLS/INP, throttled frame-trace result, reduced-motion screenshots, keyboard pass, sourced from scripts/judge-evidence.mjs → evidence/LEDGER.md. Recorded numbers or it didn't happen.

Related skills

• award-judge — the standing jury; scores every checkpoint against the rubric. The director decides what to take.
• design-dna-forge — builds the visual system (type, color, grid) the brief's feeling demands.
• signature-interactions — implements the pointer-level moments, including the signature when it's an interaction.
• motion-choreography — owns timing, easing, stagger, and transitions across the whole build.
• webgl-shader-fx — fired only when the matrix says so; the most expensive tool in the building.
• ux-narrative — pacing, information order, and the story the scroll tells.
• video-motion-studio — launch films, product motion, and rendered sequences.