All out ux

All-Out UX

The goal is positive, not the absence of failure: one committed point of view, executed to studio finish, engineering a specific reaction in the first second. Name the feeling — awe, unease, calm authority, mischief — and make every choice serve it; if a cold viewer feels nothing nameable, the pass failed no matter what it avoided. Two failure modes guard the road. Collapse inward and you get distributional convergence: the statistical center of everything you trained on, the most-seen design on earth. Overcorrect and you get costume — a novelty so desperate to be unlike anything that it reads as a gimmick instead of a resolved design. And the walls are not a corridor to walk down the middle of: equidistant from both sits "tasteful, current, safe-but-nice," which is the center again. "Would most models produce this?" kills the center; "would a top studio be embarrassed to ship this?" kills the costume; "what does it make me feel in the first second?" is the question the other two exist to serve. Past both walls hides a third failure: the showpiece — distinctive, resolved, and tiring. The reaction engineered for the first second must survive the thousandth minute; you are designing a room to work in, not a poster to glance at. Run the whole pass autonomously; state assumptions inline; never present a mode menu.

Invariants

• Grounded, not unmoored. Before diverging, read the ground truth that exists: product purpose and personas, platform conventions, repo design constraints, and prior direction verdicts. Hard constraints outrank the pass — a repo that pins shell, density, or typography is a wall, not a dare: scope the boldness to what is actually open and report the conflict in the handoff rather than bulldoze it, and never rewrite a constraint or design-language doc to legitimize the output. The committed reaction serves the product's purpose — the feeling this product needs, not the most impressive one available; awe is the wrong gift for a calm-relief tool.
• Default death test. A choice you would make on most briefs is a default, whatever it is. Steering away from defaults mints new ones — second-order convergence: the banned purple gradient of one year becomes the cream-and-terracotta cliché of the next; the anti-Inter font becomes the new Inter. Kill by the test, not by a list — any named cliché list is dated the day it's written. And distrust the test's own instrument: your sense of the center is a generation stale — trained on last year's outputs — so it under-kills this year's defaults. A direction that feels "tasteful, current, safe-but-nice" is usually the new center wearing this season's palette.
• Register, not costume. The death test's mirror image. A direction is a point of view — a tonal register carried by type, color, space, and motion — not a real-world object the interface dresses up as. The tell: if the only way to name your direction is the thing it imitates (a departure board, a control-tower strip, a carbon-copy form), you've built a costume — skeuomorphic novelty that photographs as a party trick and falls apart on use. Costume is maximally unlikely, trivially nameable, and still slop: it clears the death test and fails the brief. Find the register under the object and express that.
• Direction before code. Commit to a named art direction justified by purpose and audience before writing any markup. If the direction could caption any brief, it isn't one.
• One signature, spent boldness. Name the single treatment this piece is remembered for — a way of setting type, a compositional move, one orchestrated entrance — and concentrate the boldness there; boldness spread evenly is beige. Remembered for a move, not a mascot: a prop standing in for the signature is costume. Over-animation is itself a generic tell.
• The judge is not the maker. Self-critique flatters its own work. Score output in a separate fresh-context evaluation, and write the rubric like a prompt — its wording leaks into the visuals it selects for.
• Wild, not broken. Aesthetics get no ceiling; affordances keep their floor: readable contrast, visible focus, a working keyboard path, an honest prefers-reduced-motion fallback. Weird is the goal; broken is slop with confidence.

The pass

1. Diverge. Draft three art directions that could not have come from the same designer — three different points of view, not three different props. For each: a name, the audience rationale, the reaction its first second engineers (a nameable feeling, not a vibe-word), 4–6 named hex colors, a type pairing, the signature treatment it would be remembered for, and the one aesthetic risk it takes. Stage it as an adversarial panel, not a solo listing: three designer personas with irreconcilable philosophies each champion one direction and attack the other two before you commit — three candidates listed by one unpartitioned mind share one taste. Spread the three across the major aesthetic axes — light/dark, warm/cold, dense/airy — except where a verdict has pinned one; three exotic-darks is one candidate wearing three names. Run both invariant tests on all three — the default death test and register-not-costume — and redraft any casualty; three object metaphors means start over from register.
2. Commit. Pick the direction whose risk the brief best justifies — not the safest one. Write the tokens down; they are now law.
3. Build past comfortable. Typography carries the furthest: no stock fonts, extreme weight contrast (100/200 against 800/900, not 400 against 600), size jumps of 3x or more. Dominant color with sharp accents, never an evenly-spread palette. When the direction justifies it, reach past CSS — WebGL, shaders, canvas systems, kinetic type — under the frontier-instrument rules in the reference: token-inherited, constraint-first, performance-budgeted. Compose asymmetrically; break the grid where the signature lives. Re-compose, don't re-skin: the layout scaffold — panes, bars, toolbars, chrome — takes the death test like any color or font, and a repainted stock shell is still the stock shell; rectangles at every scale with hard seams isn't a layout, it's the wireframe left showing. Push each dimension one notch past where it feels safe — your sense of "safe" is calibrated to the center you're escaping. The direction must reach every surface: inventory the screens, overlays, footers, and empty/loading/error/ended states before building, and apply the committed tokens to all of it — a component still wearing framework-default styling (a stock indigo primary, an unthemed dialog) means the direction was never committed.
4. Judge adversarially. Evaluate with fresh context — a subagent where available, otherwise a separate pass — seeing the brief, the committed spec (direction, tokens, intended reaction), the ground truth (purpose, constraints, verdicts), and the rendered result, and judging what the pixels deliver against that spec, never the maker's narrative or effort. Two automatic fails, not one: anything the default death test catches (too central) and anything that reads as costume or as unresolved craft (novelty without finish). Originality and resolution are co-heaviest — an awful original doesn't win on nerve, it just loses. The usability floor stays pass/fail. The judge writes a per-criterion scorecard into the handoff — a pass with no scorecard is a pass that never happened. Revise and re-judge. Hard cap: five rounds or a clean verdict, whichever comes first.
5. Verify live. Real browser, motion at full speed, then the reduced-motion and keyboard paths. Walk every inventoried state — anything still wearing pre-pass or framework-default styling is an automatic fail; fix and re-verify. A screenshot proves layout, not motion. Hand off with the committed direction and signature element named, the judge scorecard, plus anything deliberately left wild.

Prior runs contribute two kinds of evidence, not just bans. A direction the user rejected votes against its axis values — dark/light, warm/cold, dense/airy — and new candidates must not cluster on a rejected value. A direction the user shipped or accepted is a verdict to respect: hold its axis values and vary the register within them, and on a bare re-invocation against an accepted incumbent, default to pushing the incumbent further (state the assumption in the handoff), reserving full re-direction for briefs that ask for it. Still ban the previous run's specific register, palette family, and typefaces from the new candidate set — repeated runs converge on this skill's own favorites too. An incumbent with no verdict behind it — template slop that merely exists — is a default like any other: ban the incumbent direction and palette family and re-direct.

Load references/creative-ceiling.md for the direction library, the frontier-instrument rules (3D, shaders, canvas, kinetic type), the judge rubric template, and the dimension-by-dimension extremity guide.