Read this first — before design-system, any register/*.md file, or any moves/*.md file. This skill owns the gate and the routing table, not the move procedures themselves.
Design Method — The Core
This is the one place the design pipeline is defined. The agent (agents/design-expert.md)
always loads this file before dispatching to a move; every references/register/*.md and
references/moves/*.md file assumes you've read this first and references it instead of
restating it — if a rule elsewhere contradicts this file, this file wins.
Setup
How this core gets loaded, every time, no exceptions:
- Read this file once per task — not once per move. Skip the re-read for a later move in
the same task if neither the Register nor the codebase changed.
- Read
../design-system/SKILL.md — always, unconditionally, before anything else
below. Thin pointer file, cheap to reload, and the harness's phase-1 trigger, so it runs
even on a task needing no new token. Existing design-system.md tokens still win over
its defaults once resolved.
- Resolve Register (below), then write the Design Read + 3 dials one-liner.
- Pass Gate 0 — Brief Lock (below) before touching any HTML/CSS/tokens.
- Read the ONE matched
references/moves/<move>.md from the Routing table — not two, not
a skim of several to "compare." The move file owns the how; this file owns the gate and
the pointer.
- If the move builds or touches code, read the matching
references/register/{brand,product}.md floors too, once.
Register
Resolve before anything else downstream — brand vs product gates every floor (Signature
Dominance, Focal-Block Floor, tone bounds, motion budget, dial presets). Never default
silently.
- brand — one dominant, expressive message: marketing site, launch page, identity.
- product — dense, predictable, motion stays discreet: dashboard, SaaS tool, utility
screen.
Priority order (fixed, first match wins):
- The owner states it explicitly.
- An existing
design-system.md/PRODUCT.md already classifies it.
- Inferable from the concrete surface (dashboard/settings/internal tool →
product;
landing/marketing/campaign/identity → brand).
- Still unclear → ask ONE question. Never guess past this point.
Once resolved, load references/register/brand.md or references/register/product.md —
each carries the floors, tone bounds and motion budget for that lane. For copy/microcopy
tone, load references/register/copy.md (any move touching user-facing text, plus
ux-copy).
Design Read + the 3 dials
Before any palette/font/layout choice, state one line:
Design Read: {page kind} for {audience}; vibe = {2-4 adjectives}; assets = {existing brand assets or "none"}; constraints = {quiet constraints or "none"}.
Then set the 3 numeric dials it implies — contractual inputs for every step after this
one, not decoration:
| Dial |
1 |
10 |
DESIGN_VARIANCE |
rigid, symmetrical, conventional |
highly art-directed, asymmetric |
VISUAL_DENSITY |
airy, gallery-like, calm |
packed, information-dense |
MOTION_INTENSITY |
static / near-still |
cinematic, scroll-driven |
Preset-by-use-case table, partial-brief fallbacks and MOTION_INTENSITY bands (calm <4 /
expressive 4–7 / cinematic >7) live in ../design-system/references/design-read-dials.md
— don't restate them here. The brief always overrides any default; an unresolvable dial
uses the documented fallback and says so in the Design Read line, never a silent guess.
What each number means — anchored on the corpus
Poles alone are unusable: nobody knows what a 7 looks like, so an unanchored dial drifts
to the middle and the middle is where every generated page already lives. The scales below
are anchored on the fifteen reference pages, each value read from that page's own
design-system.md under ../design-web/references/refs-design/ (short names below are those
directories). These are not scores. All fifteen pages were judged good; the numbers say what
kind of page it is, never how well it was done.
DESIGN_VARIANCE — how far the composition departs from the canonical skeleton.
| n |
Page |
What it looks like |
| 3 |
fora-recode |
"Bookended Canon" — nav, hero, intro, features, what-you-get, pricing, footer, almost to the letter. Reads specific anyway: one device, a 1px rule at white 10%, is the only line in the system and carries every edge, card outline and divider. |
| 3 |
cursor-recode |
"Ramp and Newsroom" — deliberately unstyled page furniture; the art direction lives entirely in the floating product window, not in the layout. |
| 3 |
endlesstools-recode |
"Showreel Canon" — the canonical order, compressed, with every slot filled by footage instead of a card. |
| 4 |
harness-recode, linear-recode, xai-recode, mosa-recode |
Canon kept, one structural inversion each: the tab row sits below the media (harness); numbered 1.0/1.1 markers turn five marketing sections into a document index (linear); the feature grid is pulled up into the hero, leaving five sections total (xai); the canonical order is followed almost exactly, but every media slot is a fake interface drawn over a photograph and features appear three times in three geometries, so there is no features block to recognise (mosa). |
| 5 |
mainframe-recode, umbrel-recode, reve-recode, parley-recode |
Skeleton rebuilt but still legible: seven short blocks under a header that scrolls away and never returns (mainframe); solo bands alternating with duet bands, then a store pivot (umbrel); seven capability panels alternating which edge they bleed off (reve); a canonical text spine cut four times by a full-bleed photographic band, and a FAQ that refuses to be an accordion — ten questions and answers all open (parley). |
| 6 |
stash-recode, stripe-recode |
Canon dissolved into a repeated device rather than reordered: six of nine sections are the same rounded photographic card and no two consecutive sections share a geometry, so the skeleton is legible only by function (stash); seventeen sections ordered by the size of the company addressed instead of by argument, each rung re-arming with the same title/action template (stripe). |
| 7 |
supercommon-recode, dispatch-recode |
"Void-Metered Stack" — testimonials, pricing table, FAQ, CTA band and any nav dropped outright; six consecutive sections share one band class and are told apart only by content; a single lowercase display line is the composition; no local asset anywhere (supercommon). Sixteen sections in which four separate surfaces recreate a running interface in HTML/CSS — two terminal chassis, a browser chassis, a chat overlay — the first pulled -6rem into the hero, and the pricing answered by an interactive calculator rather than a table (dispatch). |
VISUAL_DENSITY — how much is on screen per unit of scroll.
| n |
Page |
What it looks like |
What is on screen — counted on that folder's index.html |
| 2 |
reve-recode |
One capability per 104svh; the gutter is a slope, not a container, and text width is never capped by a wrapper. |
51 <img>, 2 <video>, 4 <source>. A 13-card drifting strip in the hero, then one full-window editing panel per capability, each carrying paired before/after frames; two of the seven are live-embed slots. Sparse vertically. Never empty. |
| 2 |
supercommon-recode |
Extremely sparse vertically, tight typographically: exactly one padding value in the whole page. |
3 photographs + 1 looping video + 18 inline SVG pictograms. The hero is a 3437×1905 image, full-bleed at sizes="100vw", under a 30vh scrim (.hero__ramp); a 2048×2048 square carries the two words; a 3200×2203 screenshot carries the last band. The one page here with no local asset still puts four remote media on screen. |
| 4 |
mainframe-recode |
Editorial: a single 128px inter-block gap repeated everywhere, so page length comes from what a block contains and never from padding — six blocks over 5,752px at 1440. |
23 <img>. One wallpaper used twice as a 140vh halo, a five-panel full-window carousel, five showcase .webp shipped twice over (two complete trees, one hidden per media query), five brand-rail logos, one 180px closing object. Every media surface is the same 10px inset-hairline frame. |
| 5 |
cursor-recode, fora-recode |
Airy between, dense within: the 24-column grid exists only inside large cards, never on the page (cursor); 36px card padding and 20px pricing rows inside airy sections (fora). |
cursor: 18 <img> (avatars and logos, 48–96px) + 20 inline SVG, 42 KB — plus a phone drawn entirely in CSS and two floating product windows on a full-bleed plate. fora: 23 <img> + 38 inline SVG — and one rendered app window carrying every claim, in the hero, then bled 272px off the CTA edge over a 190px horizon strip. |
| 6 |
umbrel-recode, stash-recode, parley-recode |
Deliberately tight — a 60px section stack at desktop that rises at narrower breakpoints, inverting the usual direction; ten sections in ~5,500px (umbrel). Nine sections in which six are one repeated rounded photographic card, each carrying its own text (stash). Fifteen sections over 10,600px, so density comes from what a band contains, never from how close the bands are (parley). |
umbrel: 42 <img> + 88 inline SVG, 75 KB — the largest drawn payload in the corpus. Full-bleed product photography on each solo band, a five-card snap rail, a collage, and two marquees running in opposite directions. stash: 6 <img> (all local, none reused) + 18 inline SVG — one photograph per photographic surface, plus a single icon sprite referenced by <use>, an agent window and a phone chassis drawn in CSS. parley: 5 <img> + 21 inline SVG — four photographs, one of them twice, four floating interface mocks and a phone chassis, all drawn. |
| 7 |
endlesstools-recode, linear-recode, xai-recode, dispatch-recode |
Split density, the mark of a high value done well: 8px gutters and 8px type inside the mosaic against 150px of nothing between sections (endlesstools); dense inside the mockups, padding-block: 128px around them (linear); dense without noise, five sections, every visual drawn in markup (xai); sixteen sections where the density sits inside four recreated interfaces and a live calculator, with a 42rem manifesto deliberately carrying no component at all (dispatch). |
endlesstools: 69 <img> + 20 <video> — a 20-tile mosaic on a fixed 1200px scene, 8 image tiles, 14 review cards. linear: 25 <img> + 95 inline SVG — six application-window mockups, each in a 22px frame with a bottom mask, plus animated dot matrices. xai: 7 <img> only (three at 320px, four news thumbs at 1200×630) + 30 inline SVG — a five-bubble conversation, a terminal with eight lines of output, a 3×2 mosaic, a waveform, a full-size code window, every one drawn in markup. dispatch: 4 <img> (local) + 10 inline SVG — and four interface surfaces built in HTML/CSS instead: two terminal chassis, a browser-chromed demo, a chat overlay laid on a photograph. |
| 8 |
mosa-recode |
Every media slot filled and none of them idle: ten sections, each one carrying either a scene, a tab-swapped pair, a snap rail or a ten-cell logo box — with one 4px radius across the entire page so the density never gets visual relief from shape. |
10 <img> from 4 local files + 0 <svg>. The four photographs are re-cropped and re-veiled per slot; every interface fragment on top of them — chat panel, chip chain, composer, avatar disc, send pill, toggle — is HTML/CSS, so the drawn payload is real but carries no SVG at all. |
| 9 |
harness-recode, stripe-recode |
Counted, not felt: a four-column mega-menu, fifteen tab panels, an eight-field form and a fifteen-link footer column (harness). The corpus ceiling, and the only page that gets there on colour as well as count: seventeen sections, seven product mockups, four customer cases, eighteen figures, two carousels and an accordion (stripe). |
harness: 91 <img> + 39 inline SVG + 2 <video> + 5 CSS url(). The customer logos are declared three times over to feed the marquee; on top of that the mega-menu, the fifteen tab panels, the form and the footer column above. stripe: 4 <img> (local) + 0 <svg> — the entire visual payload is HTML/CSS: seven product mockups, a 46-ray burst, an orchestration diagram, a dashboard panel, four floating code bubbles, a CSS-drawn book cover. The four photographs appear only inside the customer-case accordion. |
Read the fourth column before you copy a number from the first. Not one of the fifteen
holds a low density with an empty screen. reve and supercommon sit at the corpus floor
and still put fifty-one images and four media files in front of the reader; stripe holds
the corpus's highest density on four images and zero <svg>, because it draws seven
product screens in markup instead of capturing them — the same trick xai runs at 7. A density value says how much scroll a given quantity of
matter is spread over. It never says how much matter there is. Copying 2 from reve
onto a page with nothing to show does not reproduce reve — it reproduces the number and
drops the page.
The screen-holding floor — checked at Gate 0, not after the build
A dial triple says how a page is composed. It says nothing about whether there is
anything to compose. Three facts together leave the reader with nothing to look at:
VISUAL_DENSITY ≤ 3, and
DESIGN_VARIANCE ≥ 6, and
- no visual asset available — no photography, no video, no product UI, no data, no map,
no document, no object.
Each is legitimate on its own. Together they are not a direction, they are an empty page:
high variance drops the canonical furniture (card grids, logo walls, metric bands,
testimonials) that would otherwise have filled the screen; low density spreads what remains
over a long scroll; and with no asset there is nothing to put in the space the other two
just cleared. The corpus contains no such page. supercommon, its closest point at
7 · 2, still ships three photographs, a video and a metallic band.
The check is countable, and it runs before code. List what will physically be on screen,
block by block: every image, video, mockup, drawn interface, chart, table, full-bleed colour
field, and every type object large enough to be a composition in itself. If that list holds
fewer than one entry per planned section, one of three things has to change, and the plan
says which:
Raise the density — bring back the furniture the variance dropped, and treat it
(per-section padding rhythm, a rule crossing section boundaries) instead of leaving it
plain. fora at variance 3 is a corpus reference precisely for this.
Manufacture the matter. Drawn matter counts, and it is the answer when the brief
supplies no photography: xai holds density 7 on seven images by building its interface
out of markup; supercommon makes a whole band out of a gradient and one 56px line;
mainframe makes two halos out of a single image cut by opposite alpha masks.
When the brief does supply usable photography, manufacturing matter instead is not a
neutral choice — it is refusing what the subject actually has, and a page that then
labels a section "photographed" over nothing but drawings contradicts itself in public.
Drawing is a way of filling a screen; it is never a way out of using the client's
images. Use them, and draw around them. And where the brief supplies none, a free-stock
image is still reachable before a drawing is — sourcing rules live in
design-web/references/photos-images.md, the source hierarchy is enforced at
design-review/references/pre-flight-checklist.md check 15.
Better still, carry both — they reinforce each other. Drawing and photography are not
competitors, and a page holding both holds better than a page holding either alone. The
division of labour is clean: the photograph carries matter, light, the real thing; the
drawing carries structure, measure, and what the eye cannot see on its own — a section, a
survey line, a scale, a before/after. Real estate and architecture are the plain case: a
plan or a cross-section set beside a photograph of the place says what neither says
alone, the photograph showing what it is like and the drawing how it is made. One limit,
and it is intention: an SVG pictogram parked next to a photo is not this — this is a
drawing that tells the reader something the photograph cannot.
Lower the variance — a canonical skeleton at least arrives with its own blocks.
What is never an answer is shipping the triple and letting body copy carry three screens on
its own. Type at scale is matter — but only when it is the declared signature and treated
as an object, the way supercommon's solitary display line opens the page, then every band
after it, then returns split in two. Type left at body scale over an empty band is not a
composition; it is a page that failed to get made.
MOTION_INTENSITY — how much of the page moves, and how far.
| n |
Page |
What it looks like |
| 3 |
reve-recode |
0 @keyframes, 4 transitions, no scroll reveal anywhere; the hero video fades up over 150ms only once frames are proven to be rendering. |
| 3 |
linear-recode |
translateY(4px) on entry, scale(.97) on press; eighteen curves declared and not one bounces. Nothing moves; everything settles. |
| 3 |
xai-recode |
182 transition triggers, 73% of which move nothing — they change a colour. Five duration tiers ruled by "the bigger the element, the slower it is". |
| 3 |
cursor-recode |
One reveal keyframe for the whole page, always declared and paused; JS only lifts the pause, so a failed script leaves everything visible. Buttons declare no transition at all, on purpose. |
| 3 |
parley-recode |
1 @keyframes, 14 transition lines, and the reveal is deliberately kept below the threshold where it would need to earn itself: the page is 10,600px of hairline-separated bands and motion is asked to do none of the separating. |
| 4 |
umbrel-recode, supercommon-recode |
Three named entrance variants plus two marquees and a scroll-snap rail (umbrel); three durations, two curves, one 60ms cascade step and no one-off value anywhere else (supercommon). |
| 5 |
mosa-recode |
Zero @keyframes and 25 transitions: three durations (200/300/700ms) with a declared hole between them, and everything that moves is transform or opacity — no width, height, top, left, margin or padding anywhere. Coverage of every state, one reveal, and no periodic animation at all. |
| 6 |
mainframe-recode |
Total coverage, minimal amplitude: 75 transitions, zero keyframes applied, two duration regimes — response at 150/200ms, reveal at 400–1000ms — with a deliberate hole between them. |
| 6 |
harness-recode, fora-recode, endlesstools-recode |
Entrances differentiated by role, with a 300ms gap so the light arrives after the object (harness); grid-template-rows: 0fr → 1fr accordions and mask-image hairlines (fora); one curve on eleven of twelve transitions plus twenty running videos (endlesstools). |
| 6 |
stash-recode, stripe-recode |
0 @keyframes in the source and 0 here, 24 transitions, a press gesture copied literally, hover adding something in exactly two places, and one staged sequence inside the agent panel (stash). 6 @keyframes and 20 transitions: three scale-only breathing loops on the hero ribbons, a counting eyebrow, a rotating quote with progress-bar tabs, and a step-scrolled rail (stripe). |
| 7 |
dispatch-recode |
The one page above the old ceiling. Total coverage, low amplitude, one exception: every block enters on opacity 0 → 1 + a 24px rise over 700ms, while the two h1 lines get a 110% translate inside an overflow: hidden mask, 900ms, cascaded 110ms apart. On top of that, three periodic animations (SCROLL tick 2.2s, terminal caret 1.1s steps(1), demo gauge 1.6s) and four scripted behaviours — hero parallax at 0.18 × scrollY, four counters over 1400ms, tab groups, and the calculator. |
Read the corpus honestly on one point: at 6, motion is coverage, not amplitude —
mainframe and harness earn a 6 because every state is covered by a documented system, not
because anything is spectacular. 7 is now the corpus ceiling, and it was 6 until
dispatch landed. Read what moved that number, because it is not what the scale's poles
suggest: dispatch earns a 7 by adding three periodic animations and four scripted
behaviours — including a genuine parallax at 0.18 × scrollY — on top of the same total,
low-amplitude coverage the 6s have. Still no page on disk plays a pinned or scrubbed
sequence. So 7 has one anchor and it is a modest one; from 8 up the scale has poles but
no anchor at all. A brief that lands at 8 is not thereby wrong — it is unanchored, and the
obligations under A high value commits you to producing something below are the only
thing holding it. Say in the plan that nothing in the corpus demonstrates the value you
claimed.
Notches nobody occupies — 1, 2, 8, 9, 10 on variance; 1, 3, 10 on density; 1, 2, 8, 9, 10
on motion — are not forbidden. Where the notch sits between two occupied ones,
interpolate and say what the in-between buys you; where it sits above everything
(variance 8+, density 10, motion 8+), there is nothing to interpolate from and the number
is a promise with no precedent — treat it accordingly. And note what the corpus does not
contain: no page is high on all three axes. Every one of the fifteen gives something up
— harness is a 9 in density but a 4 in variance; supercommon is a 7 in variance but a 2 in
density; stripe reaches 9 in density and pays for it at 6 · 6 on the other two; reve sits
at the corpus floor on two axes at once, 5 · 2 · 3, and is the only page that does. The
closest thing to an exception is dispatch at 7 · 7 · 7, and it is worth naming as
such: it is the only page that refuses to give an axis up, and the price is visible — it
carries the corpus's most complex motion budget on a page that also has to keep four
recreated interfaces legible. A brief that arrives at 8/8/8 has not made a choice, it has
made a wish.
One more thing the comparison exposes, and it now cuts both ways. Where a preset row sits
below what the subject supports, the corpus proves it: the "SaaS / product landing"
default is 4 · 5 · 4 where harness-recode, a SaaS landing, ships 4 · 9 · 6 — and no
preset row anywhere reaches the 9 that harness holds. But two rows sit above everything
on disk — agency 8 · 4 · 7 and luxury 5 · 3 · 7 both claim a variance or a motion value no
page here occupies. Those two are unanchored in this corpus, so a brief landing on one of
them cannot check its number against anything. Either way the rule is the same: the
presets are a safe starting point for a thin brief, never a target, and the dial is argued
from the subject.
Setting a dial is a decision, not a field
Each dial is argued from the subject in the same breath as the register and the tone,
appended to the Design Read line:
Dials: DESIGN_VARIANCE n · VISUAL_DENSITY n · MOTION_INTENSITY n — because {what in the subject demands it}.
A law firm at DESIGN_VARIANCE 8 owes exactly the sentence a fashion label at 3 owes.
Neither number is wrong; an unargued one always is.
A low value is not a low score. fora at variance 3, cursor at 3, reve at density
2 — all judged good. What separates a chosen low value from a timid one is that it is
executed to the end: fora commits to a single 1px rule and uses nothing else; reve gives one
capability the whole screen and never crowds it. A low value assumed and carried through
beats a high value claimed and not delivered, every time.
The middle is suspect. Three 5s across three axes is the signature of a decision not
made. It is not banned — but it is worth knowing that no page in the corpus carries a 5 on
more than one axis. The two pages closest to the middle each buy it with a named device:
umbrel (5 · 6 · 4) with its solo/duet alternation, mainframe (5 · 4 · 6) with a brand
page that argues in the product register. So: if two or more dials land at 5, the plan says
in one line why the middle is the answer here, or moves one of them. Flag it as such — a
silent 5/5/5 is the failure this rule exists to catch.
A high value commits you to producing something. The number is a promise the deliverable
has to keep:
DESIGN_VARIANCE ≥ 7 obliges a macrostructure that visibly departs from the canon and
a signature procedure derived from the subject, applied more than once. supercommon is
the only page here that clears 7, and it carries both: its solitary display line opens the
page and then every band after it, then returns split in two as stillness. / motion.;
and its voids are metered — each silence sized to the block behind it, the two largest
(32vh, 40vh) reserved for the two blocks the page most wants read. Without both, high
variance is novelty, not direction.
VISUAL_DENSITY ≥ 7 obliges a counted inventory of what is on screen and two spacing
regimes — dense inside blocks, airy between them. Every high-density page in the corpus
splits this way. Density without the second regime is noise.
VISUAL_DENSITY ≤ 3 obliges the same counted inventory, for the opposite reason: it
has to show that something still holds each screen. See The screen-holding floor above,
and the fourth column of the density table — both corpus pages at 2 are carrying real
matter. A low density with nothing counted is not restraint, it is an empty page.
MOTION_INTENSITY ≥ 6 obliges a documented motion system: the material list, the duration
set, the curve set, and entrances differentiated by role rather than one blanket reveal.
Above 7, add a named scroll mechanism and its prefers-reduced-motion path.
If the deliverable does not carry the obligation, lower the dial in the plan rather than
ship the number — a claimed 8 that the page does not hold is the more expensive
mistake, because every downstream move inherits it.
The macrostructure pick and DESIGN_VARIANCE must agree. A canonical skeleton is a 3; a
rebuilt-but-legible one a 5; dropped sections and a bespoke order a 7, which is where the
corpus stops — no page on disk shows what an 8 or a 9 costs to hold. A plan naming the
canonical skeleton next to DESIGN_VARIANCE 8 has one of the two wrong — resolve it before
Gate 0, not while coding.
Gate 0 — Brief Lock
Before routing to any target skill — before writing or modifying a single line of
HTML/CSS/tokens — six artefacts must exist in writing, not just in your head:
Register — resolved above, stated explicitly (not defaulted).
Tone committed to ONE extreme — not an adjective that could also describe three
competitors. Where it comes from depends on whether tone is one of the axes the
Exploration Gate put in play:
- Explored — the gate fans out 3 divergent, text-only direction sketches judged
blind and comparatively by
fuse-ai-pilot:challenger, and the winner's tone becomes
this artefact. The gate runs before Step 1 of generate wherever that move has a
page body to decide — it is no longer a single scope/register special case.
- Not explored — either no gate ran, or it ran on a reduced axis list because tone
was already arbitrated in writing. Tone is then stated directly:
generate derives it
from the brief, refinement moves inherit it from the existing surface and name it.
Inheriting is not skipping — an inherited tone left unstated fails this gate exactly
like a missing one.
Which case applies, and which axes each scope × register combination puts in play, is
defined once, in the scope table of references/moves/generate.md (§Exploration
Gate). Read it there. It is deliberately not restated here — two copies of one rule
are two rules that drift apart, and that file also owns the mechanics, the "reduced is
not lighter" rule, and the fallback.
The fallback survives unchanged and is the only route back to an unexplored direction
when the gate was due: Task/Agent unavailable at nesting depth 5 → the direct single
pick, and the report marked "direction not explored / single-fiat". Never a silent one.
Signature element (brand register) or primary task (product register) named in
one sentence.
The corpus actually seen, then read — in that order, and both halves are required:
- at least two of the fifteen pages opened in a browser and scrolled through, named in
the plan as
{reference}-recode/index.html. They open by double-click, file://, no
server, no build;
design-web/references/refs-design/README.md plus at least two tokens-*.md
sections, cited by reference and section name.
This is the evidence artefact, per design-web/references/design-inspiration.md.
Sections cited with no page opened does not satisfy it. Measured values let you
reproduce a procedure, never choose one, and a corpus of taste that is only ever read
hands back conformity instead of judgment — the failure is documented at
design-web/references/refs-design/README.md §Look before you read. A sector URL or a
supplied screenshot is an addition to this artefact, never a substitute for either half.
What will be on screen, counted — one list, block by block, of the images, videos,
mockups, drawn interfaces, charts, tables, full-bleed colour fields and display-scale
type objects the page will actually carry. Required unconditionally, and decisive
whenever VISUAL_DENSITY ≤ 3 or DESIGN_VARIANCE ≥ 6: see The screen-holding floor
above for what to do when the count comes up short. A plan that names a dial triple but
cannot say what fills the screen has not been made.
Structure named, both halves — one line in the plan:
Macrostructure: {hero treatment from references/macrostructure-bank.md} + {body sequence from references/body-sequence-bank.md}
Naming one without the other fails this gate. The two banks are separate on purpose:
the hero bank owns the first screen only, the body-sequence bank owns everything
after it. The body, not the hero, is the documented failure mode
(references/register/brand.md intro) — a plan that names a hero treatment and stops
has improvised the scroll, and improvised scrolls converge on the same
feature-grid-then-testimonials default the corpus was assembled to break. So an unnamed
body sequence is a missing artefact, not a detail to settle while coding.
Restate the sequence's ordering principle in the brief's own terms, not the bank's
label: "four chambers, one image transformation each" is a body sequence, "alternating"
is a word. And the pick has to agree with DESIGN_VARIANCE (see the dial section
above) — a canonical skeleton and a variance of 8 cannot both be true.
Where the Exploration Gate ran, this pair is its output, not a fresh pick: the
winning sketch's hero treatment + body sequence, carried forward with its principle
line. Recording it satisfies this artefact; re-deciding it discards the comparison that
was just paid for. Where no gate ran, pick both from the banks here.
This is a present/absent check on six named artefacts — structural, not a taste
judgment. One missing ⇒ generation or modification is forbidden until it exists.
Mechanics of artefacts 1–2 (the Exploration Gate and its scope table, brief questions,
signature element, two-pass critique) live in references/moves/generate.md; the two banks
own the vocabulary of artefact 5. This section is only the checkpoint every move respects
before it's allowed to touch code — it says what must exist, never how it was produced.
Design guidance
Dense and general, applying whichever move is running — taste rules specific to one move
live in that move's file.
Absolute bans
Match-and-refuse, unconditional across both registers — not a statistical reflex to weigh,
a hard stop. If you're about to write any of these, rewrite the element with a different
structure, never a variation on the same idea.
Side-stripe borders — border-left/border-right >1px as a colored accent on cards,
list items, callouts, or alerts. Rewrite with full borders, background tints, leading
numbers/icons, or nothing.
Gradient text — background-clip: text combined with a gradient background.
Decorative, never meaningful. Single solid color; emphasis via weight or size.
Glassmorphism as default — blurs/glass cards used decoratively rather than rare and
purposeful, or nothing.
The hero-metric template — big number, small label, supporting stats, gradient
accent. SaaS cliché; see register/brand.md §2 "Ribbon 4-stats" furniture for the fix.
Identical card grids — same-sized cards, icon + heading + text, repeated endlessly.
Modal as first thought — modals are usually laziness. Exhaust inline/progressive
alternatives first.
Em dash in copy — commas, colons, semicolons, periods, or parentheses instead. Also
not --.
The AI-slop signature triad — defined here and canonical here; every other file in
the pipeline points back to this bullet instead of restating it. Three tells that each
pass on their own and together fingerprint the page:
- a Tailwind blue/indigo primary or accent, hue range 200–290°;
- Inter as the primary typeface;
rounded-2xl as the default radius.
That palette/font/radius combination is shared by ~83% of sampled AI-generated pages
(grounding: sailop ai-slop research). The figure belongs to this triad and to nothing
else — never quote it for another ban, another detector, or a general anti-slop claim.
Any two legs are survivable; the third is what makes the page identifiable. So if brand
guidelines already fix 2 of the 3, the third MUST change — another hue, another radius,
or another typeface (reflex-reject list and alternatives:
design-system/references/forbidden-fonts.md). The corpus carries the counter-example:
fora-recode and umbrel-recode both ship Inter and both read specific, because neither
carries either of the other two legs.
Deterministic grep companions for the structural ones (side-stripe, identical grids,
nested cards, icon-bento): design-review/references/anti-ai-slop-audit.md.
The AI-slop test — 2 levels
If someone could look at this interface and say "AI made that" without doubt, it's failed.
Run at two altitudes — the second catches what the first misses. Cross-register
failures are the Absolute bans above; register-specific failures live in each
register/*.md.
First-order — category reflex. If someone could guess the theme + palette from the
category alone, it's the first training-data reflex:
- Cream #F4F1EA + a contrasted serif + terracotta accent — the default "editorial
SaaS" look.
- Near-black background + one acid accent color — the default "dark developer tool"
look.
- Broadsheet hairlines, zero border-radius, black/white only — the default "premium
minimal" look.
- Glassmorphism +
rounded-2xl used globally — the default "2026 AI app" look,
applied everywhere instead of gated (design-motion gates it deliberately).
- Generic icon-bento — every cell centered text over a round colored-icon badge, zero
image/gradient/pattern variation between cells.
Rework the Design Read's vibe words and the color strategy (design-system §Color
strategy) until the answer isn't obvious from the category — "observability → dark blue",
"healthcare → white + teal", "finance → navy + gold", "crypto → neon on black" are all this
tier. Purple-on-white gradients are banned outright regardless of tier — the single most
common tell.
Second-order — category-plus-anti-reference trap. The first reflex was avoided, but if
someone could still guess the aesthetic family from category-plus-anti-references — "AI
workflow tool that's not SaaS-cream → editorial-typographic", "fintech that's not
navy-and-gold → terminal-native dark mode" — it's the trap one tier deeper. Rework until
neither answer is obvious. Register-specific reflex-reject aesthetic lanes:
register/brand.md and register/product.md.
Deterministic grep detectors cover the first-order clusters plus the Absolute bans; the
second-order trap is judgment, checked at the challenger gate
(design-review/references/review-procedure.md), not grep-lintable.
Macrostructure variety — hero AND body
Centered hero + 3-column icon-card grid is forbidden as a default skeleton. Before any plan
is finalized, name two picks: a hero treatment from references/macrostructure-bank.md
(first screen only) and a body sequence from references/body-sequence-bank.md (everything
after it, with its ordering principle restated in the brief's own terms). A plan carrying
only a hero treatment has improvised the scroll — the body is the documented failure mode.
Following the canon is a legitimate pick. fora-recode's "Bookended Canon" runs the
canonical skeleton almost to the letter and does not read generic, because one device — a
single 1px rule at white 10% — is carried through every edge, card and divider.
endlesstools-recode keeps the same order and fills every slot with footage. Departing is
equally legitimate: supercommon-recode's "Void-Metered Stack" drops the testimonials, the
pricing table, the FAQ, the CTA band and the nav outright — the price is inline on one
button and the exit is a single link with no headline above it. What is never legitimate is
not choosing — the skeleton that arrived by default is the one that reads as generated.
Let each pick follow the subject, not habit: a treatment or a sequence reused by reflex
across briefs becomes its own template, whichever one it is. A pick used on two consecutive
briefs has to be argued from the second subject or replaced — the fact that it worked once
is not an argument.
Non-negotiable floor
Regardless of tone, register, target, or move, every deliverable must have: responsive
behavior across the target's breakpoints/size classes; visible keyboard focus on every
interactive element (:focus-visible, never suppressed); prefers-reduced-motion
respected wherever motion is added.
Generation approach
Generate HTML/CSS directly — the default and primary path, following the same method as
Anthropic's official frontend-design skill (commit to a point of view, avoid templated
defaults, verify with tools not vibes). Gemini
…(truncated)
1---2name: design-method3description: "Use when starting any design task — resolves register, sets the 3 dials, locks Gate 0, and routes to the right move and target skill."4when-to-use: "Start of any design task, before picking a move or writing/editing a single line of HTML/CSS/tokens."5keywords: register, brand, product, dials, gate-0, absolute-bans, anti-slop, routing, moves, method6priority: critical7related: design-system, design-web, design-webapp, design-ios, design-android, design-review, ux-copy8metadata:9 source: https://skillmd.com/skills/fusengine/design-method10---1112<objective>13The core design method and entry point for any design task: resolves the Register (brand vs product), states the Design Read one-liner and sets the 3 numeric dials (DESIGN_VARIANCE, VISUAL_DENSITY, MOTION_INTENSITY) against a scale anchored on the fifteen reference pages, enforces the Gate 0 brief lock before any HTML/CSS/tokens are touched, applies the Absolute bans, runs the 2-level AI-slop test, and routes to the matching move (generate, critique, audit, bolder, quieter, distill, harden, polish, redesign) and its target skill.1415Read this first — before `design-system`, any `register/*.md` file, or any `moves/*.md` file. This skill owns the gate and the routing table, not the move procedures themselves.16</objective>1718<!-- Grounding: grounding-corpus.md §A (impeccable anatomy — thin router, Register, dials, Absolute bans, 2-level slop test), §G (root-cause fix: template reproduction demoted, register-first), §H (kept anchors: Gate 0, Signature Dominance, Focal-Block Floor, Competitor Lift Test, MOTION_INTENSITY). -->1920# Design Method — The Core2122This is the one place the design pipeline is defined. The agent (`agents/design-expert.md`)23always loads this file before dispatching to a move; every `references/register/*.md` and24`references/moves/*.md` file assumes you've read this first and references it instead of25restating it — if a rule elsewhere contradicts this file, this file wins.2627## Setup2829How this core gets loaded, every time, no exceptions:30311. Read this file once per task — not once per move. Skip the re-read for a later move in32 the same task if neither the Register nor the codebase changed.332. **Read `../design-system/SKILL.md`** — always, unconditionally, before anything else34 below. Thin pointer file, cheap to reload, and the harness's phase-1 trigger, so it runs35 even on a task needing no new token. Existing `design-system.md` tokens still win over36 its defaults once resolved.373. Resolve **Register** (below), then write the **Design Read + 3 dials** one-liner.384. Pass **Gate 0 — Brief Lock** (below) before touching any HTML/CSS/tokens.395. Read the ONE matched `references/moves/<move>.md` from the Routing table — not two, not40 a skim of several to "compare." The move file owns the how; this file owns the gate and41 the pointer.426. If the move builds or touches code, read the matching43 `references/register/{brand,product}.md` floors too, once.4445## Register4647Resolve **before anything else downstream** — brand vs product gates every floor (Signature48Dominance, Focal-Block Floor, tone bounds, motion budget, dial presets). Never default49silently.5051- **brand** — one dominant, expressive message: marketing site, launch page, identity.52- **product** — dense, predictable, motion stays discreet: dashboard, SaaS tool, utility53 screen.5455**Priority order (fixed, first match wins):**561. The owner states it explicitly.572. An existing `design-system.md`/`PRODUCT.md` already classifies it.583. Inferable from the concrete surface (dashboard/settings/internal tool → `product`;59 landing/marketing/campaign/identity → `brand`).604. Still unclear → ask **ONE** question. Never guess past this point.6162Once resolved, load `references/register/brand.md` or `references/register/product.md` —63each carries the floors, tone bounds and motion budget for that lane. For copy/microcopy64tone, load `references/register/copy.md` (any move touching user-facing text, plus65`ux-copy`).6667## Design Read + the 3 dials6869Before any palette/font/layout choice, state one line:7071```72Design Read: {page kind} for {audience}; vibe = {2-4 adjectives}; assets = {existing brand assets or "none"}; constraints = {quiet constraints or "none"}.73```7475Then set the 3 numeric dials it implies — **contractual inputs** for every step after this76one, not decoration:7778| Dial | 1 | 10 |79|------|---|----|80| `DESIGN_VARIANCE` | rigid, symmetrical, conventional | highly art-directed, asymmetric |81| `VISUAL_DENSITY` | airy, gallery-like, calm | packed, information-dense |82| `MOTION_INTENSITY` | static / near-still | cinematic, scroll-driven |8384Preset-by-use-case table, partial-brief fallbacks and `MOTION_INTENSITY` bands (calm <4 /85expressive 4–7 / cinematic >7) live in `../design-system/references/design-read-dials.md`86— don't restate them here. The brief always overrides any default; an unresolvable dial87uses the documented fallback and says so in the Design Read line, never a silent guess.8889### What each number means — anchored on the corpus9091Poles alone are unusable: nobody knows what a 7 looks like, so an unanchored dial drifts92to the middle and the middle is where every generated page already lives. The scales below93are anchored on the fifteen reference pages, each value read from that page's own94`design-system.md` under `../design-web/references/refs-design/` (short names below are those95directories). These are not scores. All fifteen pages were judged good; the numbers say what96kind of page it is, never how well it was done.9798**`DESIGN_VARIANCE` — how far the composition departs from the canonical skeleton.**99100| n | Page | What it looks like |101|---|---|---|102| 3 | `fora-recode` | "Bookended Canon" — nav, hero, intro, features, what-you-get, pricing, footer, almost to the letter. Reads specific anyway: one device, a 1px rule at white 10%, is the only line in the system and carries every edge, card outline and divider. |103| 3 | `cursor-recode` | "Ramp and Newsroom" — deliberately unstyled page furniture; the art direction lives entirely in the floating product window, not in the layout. |104| 3 | `endlesstools-recode` | "Showreel Canon" — the canonical order, compressed, with every slot filled by footage instead of a card. |105| 4 | `harness-recode`, `linear-recode`, `xai-recode`, `mosa-recode` | Canon kept, one structural inversion each: the tab row sits *below* the media (harness); numbered `1.0`/`1.1` markers turn five marketing sections into a document index (linear); the feature grid is pulled up *into* the hero, leaving five sections total (xai); the canonical order is followed almost exactly, but every media slot is a fake interface drawn over a photograph and features appear three times in three geometries, so there is no features block to recognise (mosa). |106| 5 | `mainframe-recode`, `umbrel-recode`, `reve-recode`, `parley-recode` | Skeleton rebuilt but still legible: seven short blocks under a header that scrolls away and never returns (mainframe); solo bands alternating with duet bands, then a store pivot (umbrel); seven capability panels alternating which edge they bleed off (reve); a canonical text spine cut four times by a full-bleed photographic band, and a FAQ that refuses to be an accordion — ten questions and answers all open (parley). |107| 6 | `stash-recode`, `stripe-recode` | Canon dissolved into a repeated device rather than reordered: six of nine sections are the same rounded photographic card and no two consecutive sections share a geometry, so the skeleton is legible only by function (stash); seventeen sections ordered by the *size of the company addressed* instead of by argument, each rung re-arming with the same title/action template (stripe). |108| 7 | `supercommon-recode`, `dispatch-recode` | "Void-Metered Stack" — testimonials, pricing table, FAQ, CTA band and any nav dropped outright; six consecutive sections share one `band` class and are told apart only by content; a single lowercase display line is the composition; no local asset anywhere (supercommon). Sixteen sections in which four separate surfaces recreate a running interface in HTML/CSS — two terminal chassis, a browser chassis, a chat overlay — the first pulled `-6rem` *into* the hero, and the pricing answered by an interactive calculator rather than a table (dispatch). |109110**`VISUAL_DENSITY` — how much is on screen per unit of scroll.**111112| n | Page | What it looks like | What is on screen — counted on that folder's `index.html` |113|---|---|---|---|114| 2 | `reve-recode` | One capability per `104svh`; the gutter is a slope, not a container, and text width is never capped by a wrapper. | **51 `<img>`, 2 `<video>`, 4 `<source>`.** A 13-card drifting strip in the hero, then one full-window editing panel per capability, each carrying paired before/after frames; two of the seven are live-embed slots. Sparse *vertically*. Never empty. |115| 2 | `supercommon-recode` | Extremely sparse vertically, tight typographically: exactly one padding value in the whole page. | **3 photographs + 1 looping video + 18 inline SVG pictograms.** The hero is a 3437×1905 image, full-bleed at `sizes="100vw"`, under a 30vh scrim (`.hero__ramp`); a 2048×2048 square carries the two words; a 3200×2203 screenshot carries the last band. The one page here with no *local* asset still puts four remote media on screen. |116| 4 | `mainframe-recode` | Editorial: a **single** 128px inter-block gap repeated everywhere, so page length comes from what a block contains and never from padding — six blocks over 5,752px at 1440. | **23 `<img>`.** One wallpaper used twice as a 140vh halo, a five-panel full-window carousel, five showcase `.webp` shipped twice over (two complete trees, one hidden per media query), five brand-rail logos, one 180px closing object. Every media surface is the same 10px inset-hairline frame. |117| 5 | `cursor-recode`, `fora-recode` | Airy between, dense within: the 24-column grid exists only *inside* large cards, never on the page (cursor); 36px card padding and 20px pricing rows inside airy sections (fora). | cursor: **18 `<img>`** (avatars and logos, 48–96px) **+ 20 inline SVG, 42 KB** — plus a phone drawn entirely in CSS and two floating product windows on a full-bleed plate. fora: **23 `<img>` + 38 inline SVG** — and *one* rendered app window carrying every claim, in the hero, then bled 272px off the CTA edge over a 190px horizon strip. |118| 6 | `umbrel-recode`, `stash-recode`, `parley-recode` | Deliberately tight — a 60px section stack at desktop that *rises* at narrower breakpoints, inverting the usual direction; ten sections in ~5,500px (umbrel). Nine sections in which six are one repeated rounded photographic card, each carrying its own text (stash). Fifteen sections over 10,600px, so density comes from *what a band contains*, never from how close the bands are (parley). | umbrel: **42 `<img>` + 88 inline SVG, 75 KB — the largest drawn payload in the corpus.** Full-bleed product photography on each solo band, a five-card snap rail, a collage, and two marquees running in opposite directions. stash: **6 `<img>` (all local, none reused) + 18 inline SVG** — one photograph per photographic surface, plus a single icon sprite referenced by `<use>`, an agent window and a phone chassis drawn in CSS. parley: **5 `<img>` + 21 inline SVG** — four photographs, one of them twice, four floating interface mocks and a phone chassis, all drawn. |119| 7 | `endlesstools-recode`, `linear-recode`, `xai-recode`, `dispatch-recode` | Split density, the mark of a high value done well: 8px gutters and 8px type inside the mosaic against 150px of nothing between sections (endlesstools); dense inside the mockups, `padding-block: 128px` around them (linear); dense without noise, five sections, every visual drawn in markup (xai); sixteen sections where the density sits inside four recreated interfaces and a live calculator, with a 42rem manifesto deliberately carrying no component at all (dispatch). | endlesstools: **69 `<img>` + 20 `<video>`** — a 20-tile mosaic on a fixed 1200px scene, 8 image tiles, 14 review cards. linear: **25 `<img>` + 95 inline SVG** — six application-window mockups, each in a 22px frame with a bottom mask, plus animated dot matrices. xai: **7 `<img>` only** (three at 320px, four news thumbs at 1200×630) **+ 30 inline SVG** — a five-bubble conversation, a terminal with eight lines of output, a 3×2 mosaic, a waveform, a full-size code window, every one drawn in markup. dispatch: **4 `<img>` (local) + 10 inline SVG** — and four interface surfaces built in HTML/CSS instead: two terminal chassis, a browser-chromed demo, a chat overlay laid on a photograph. |120| 8 | `mosa-recode` | Every media slot filled and none of them idle: ten sections, each one carrying either a scene, a tab-swapped pair, a snap rail or a ten-cell logo box — with one 4px radius across the entire page so the density never gets visual relief from shape. | **10 `<img>` from 4 local files + 0 `<svg>`.** The four photographs are re-cropped and re-veiled per slot; every interface fragment on top of them — chat panel, chip chain, composer, avatar disc, send pill, toggle — is HTML/CSS, so the drawn payload is real but carries no SVG at all. |121| 9 | `harness-recode`, `stripe-recode` | Counted, not felt: a four-column mega-menu, fifteen tab panels, an eight-field form and a fifteen-link footer column (harness). The corpus ceiling, and the only page that gets there on colour as well as count: seventeen sections, seven product mockups, four customer cases, eighteen figures, two carousels and an accordion (stripe). | harness: **91 `<img>` + 39 inline SVG + 2 `<video>` + 5 CSS `url()`.** The customer logos are declared three times over to feed the marquee; on top of that the mega-menu, the fifteen tab panels, the form and the footer column above. stripe: **4 `<img>` (local) + 0 `<svg>`** — the entire visual payload is HTML/CSS: seven product mockups, a 46-ray burst, an orchestration diagram, a dashboard panel, four floating code bubbles, a CSS-drawn book cover. The four photographs appear only inside the customer-case accordion. |122123**Read the fourth column before you copy a number from the first.** Not one of the fifteen124holds a low density with an empty screen. `reve` and `supercommon` sit at the corpus floor125and still put fifty-one images and four media files in front of the reader; `stripe` holds126the corpus's *highest* density on **four** images and zero `<svg>`, because it draws seven127product screens in markup instead of capturing them — the same trick `xai` runs at 7. A density value says how much scroll a given quantity of128matter is spread over. **It never says how much matter there is.** Copying `2` from `reve`129onto a page with nothing to show does not reproduce `reve` — it reproduces the number and130drops the page.131132#### The screen-holding floor — checked at Gate 0, not after the build133134A dial triple says how a page is *composed*. It says nothing about whether there is135anything to compose. Three facts together leave the reader with nothing to look at:136137- `VISUAL_DENSITY` ≤ 3, **and**138- `DESIGN_VARIANCE` ≥ 6, **and**139- no visual asset available — no photography, no video, no product UI, no data, no map,140 no document, no object.141142Each is legitimate on its own. Together they are not a direction, they are an empty page:143high variance drops the canonical furniture (card grids, logo walls, metric bands,144testimonials) that would otherwise have filled the screen; low density spreads what remains145over a long scroll; and with no asset there is nothing to put in the space the other two146just cleared. **The corpus contains no such page.** `supercommon`, its closest point at1477 · 2, still ships three photographs, a video and a metallic band.148149The check is **countable, and it runs before code**. List what will physically be on screen,150block by block: every image, video, mockup, drawn interface, chart, table, full-bleed colour151field, and every type object large enough to be a composition in itself. If that list holds152fewer than one entry per planned section, one of three things has to change, and the plan153says which:1541551. **Raise the density** — bring back the furniture the variance dropped, and *treat* it156 (per-section padding rhythm, a rule crossing section boundaries) instead of leaving it157 plain. `fora` at variance 3 is a corpus reference precisely for this.1582. **Manufacture the matter.** Drawn matter counts, and it is the answer when the brief159 supplies no photography: `xai` holds density 7 on seven images by building its interface160 out of markup; `supercommon` makes a whole band out of a gradient and one 56px line;161 `mainframe` makes two halos out of a single image cut by opposite alpha masks.162163 **When the brief does supply usable photography, manufacturing matter instead is not a164 neutral choice** — it is refusing what the subject actually has, and a page that then165 labels a section "photographed" over nothing but drawings contradicts itself in public.166 Drawing is a way of filling a screen; it is never a way out of using the client's167 images. Use them, and draw around them. And where the brief supplies none, a free-stock168 image is still reachable before a drawing is — sourcing rules live in169 `design-web/references/photos-images.md`, the source hierarchy is enforced at170 `design-review/references/pre-flight-checklist.md` check 15.171172 **Better still, carry both — they reinforce each other.** Drawing and photography are not173 competitors, and a page holding both holds better than a page holding either alone. The174 division of labour is clean: the photograph carries matter, light, the real thing; the175 drawing carries structure, measure, and what the eye cannot see on its own — a section, a176 survey line, a scale, a before/after. Real estate and architecture are the plain case: a177 plan or a cross-section set beside a photograph of the place says what neither says178 alone, the photograph showing what it is like and the drawing how it is made. One limit,179 and it is intention: an SVG pictogram parked next to a photo is not this — this is a180 drawing that tells the reader something the photograph cannot.1813. **Lower the variance** — a canonical skeleton at least arrives with its own blocks.182183What is never an answer is shipping the triple and letting body copy carry three screens on184its own. Type at scale *is* matter — but only when it is the declared signature and treated185as an object, the way `supercommon`'s solitary display line opens the page, then every band186after it, then returns split in two. Type left at body scale over an empty band is not a187composition; it is a page that failed to get made.188189**`MOTION_INTENSITY` — how much of the page moves, and how far.**190191| n | Page | What it looks like |192|---|---|---|193| 3 | `reve-recode` | 0 `@keyframes`, 4 transitions, no scroll reveal anywhere; the hero video fades up over 150ms only once frames are proven to be rendering. |194| 3 | `linear-recode` | `translateY(4px)` on entry, `scale(.97)` on press; eighteen curves declared and not one bounces. Nothing moves; everything settles. |195| 3 | `xai-recode` | 182 transition triggers, 73% of which move nothing — they change a colour. Five duration tiers ruled by "the bigger the element, the slower it is". |196| 3 | `cursor-recode` | One reveal keyframe for the whole page, always declared and *paused*; JS only lifts the pause, so a failed script leaves everything visible. Buttons declare no transition at all, on purpose. |197| 3 | `parley-recode` | 1 `@keyframes`, 14 transition lines, and the reveal is deliberately kept below the threshold where it would need to earn itself: the page is 10,600px of hairline-separated bands and motion is asked to do none of the separating. |198| 4 | `umbrel-recode`, `supercommon-recode` | Three named entrance variants plus two marquees and a `scroll-snap` rail (umbrel); three durations, two curves, one 60ms cascade step and no one-off value anywhere else (supercommon). |199| 5 | `mosa-recode` | **Zero `@keyframes`** and 25 transitions: three durations (200/300/700ms) with a declared hole between them, and everything that moves is `transform` or `opacity` — no `width`, `height`, `top`, `left`, `margin` or `padding` anywhere. Coverage of every state, one reveal, and no periodic animation at all. |200| 6 | `mainframe-recode` | Total coverage, minimal amplitude: 75 transitions, zero keyframes applied, two duration regimes — response at 150/200ms, reveal at 400–1000ms — with a deliberate hole between them. |201| 6 | `harness-recode`, `fora-recode`, `endlesstools-recode` | Entrances differentiated by role, with a 300ms gap so the light arrives after the object (harness); `grid-template-rows: 0fr → 1fr` accordions and `mask-image` hairlines (fora); one curve on eleven of twelve transitions plus twenty running videos (endlesstools). |202| 6 | `stash-recode`, `stripe-recode` | 0 `@keyframes` in the source and 0 here, 24 transitions, a press gesture copied literally, hover adding something in exactly two places, and one staged sequence inside the agent panel (stash). 6 `@keyframes` and 20 transitions: three `scale`-only breathing loops on the hero ribbons, a counting eyebrow, a rotating quote with progress-bar tabs, and a step-scrolled rail (stripe). |203| 7 | `dispatch-recode` | **The one page above the old ceiling.** Total coverage, low amplitude, one exception: every block enters on `opacity 0 → 1` + a 24px rise over 700ms, while the two h1 lines get a 110% translate inside an `overflow: hidden` mask, 900ms, cascaded 110ms apart. On top of that, three periodic animations (SCROLL tick 2.2s, terminal caret 1.1s `steps(1)`, demo gauge 1.6s) and four scripted behaviours — hero parallax at `0.18 × scrollY`, four counters over 1400ms, tab groups, and the calculator. |204205Read the corpus honestly on one point: at 6, motion is **coverage**, not amplitude —206mainframe and harness earn a 6 because every state is covered by a documented system, not207because anything is spectacular. **7 is now the corpus ceiling, and it was 6 until208`dispatch` landed.** Read what moved that number, because it is not what the scale's poles209suggest: `dispatch` earns a 7 by adding three *periodic* animations and four *scripted*210behaviours — including a genuine parallax at `0.18 × scrollY` — on top of the same total,211low-amplitude coverage the 6s have. Still **no page on disk plays a pinned or scrubbed212sequence.** So 7 has one anchor and it is a modest one; from 8 up the scale has poles but213no anchor at all. A brief that lands at 8 is not thereby wrong — it is unanchored, and the214obligations under *A high value commits you to producing something* below are the only215thing holding it. Say in the plan that nothing in the corpus demonstrates the value you216claimed.217218Notches nobody occupies — 1, 2, 8, 9, 10 on variance; 1, 3, 10 on density; 1, 2, 8, 9, 10219on motion — are not forbidden. Where the notch sits *between* two occupied ones,220interpolate and say what the in-between buys you; where it sits *above* everything221(variance 8+, density 10, motion 8+), there is nothing to interpolate from and the number222is a promise with no precedent — treat it accordingly. And note what the corpus does *not*223contain: **no page is high on all three axes.** Every one of the fifteen gives something up224— harness is a 9 in density but a 4 in variance; supercommon is a 7 in variance but a 2 in225density; stripe reaches 9 in density and pays for it at 6 · 6 on the other two; reve sits226at the corpus floor on two axes at once, 5 · 2 · 3, and is the only page that does. The227closest thing to an exception is `dispatch` at **7 · 7 · 7**, and it is worth naming as228such: it is the only page that refuses to give an axis up, and the price is visible — it229carries the corpus's most complex motion budget on a page that also has to keep four230recreated interfaces legible. A brief that arrives at 8/8/8 has not made a choice, it has231made a wish.232233One more thing the comparison exposes, and it now cuts both ways. Where a preset row sits234*below* what the subject supports, the corpus proves it: the "SaaS / product landing"235default is 4 · 5 · 4 where `harness-recode`, a SaaS landing, ships 4 · 9 · 6 — and no236preset row anywhere reaches the 9 that harness holds. But two rows sit *above* everything237on disk — agency 8 · 4 · 7 and luxury 5 · 3 · 7 both claim a variance or a motion value no238page here occupies. Those two are unanchored in this corpus, so a brief landing on one of239them cannot check its number against anything. Either way the rule is the same: the240presets are a safe starting point for a thin brief, never a target, and the dial is argued241from the subject.242243### Setting a dial is a decision, not a field244245Each dial is argued from the subject in the same breath as the register and the tone,246appended to the Design Read line:247248```249Dials: DESIGN_VARIANCE n · VISUAL_DENSITY n · MOTION_INTENSITY n — because {what in the subject demands it}.250```251252A law firm at `DESIGN_VARIANCE 8` owes exactly the sentence a fashion label at 3 owes.253Neither number is wrong; an unargued one always is.254255**A low value is not a low score.** `fora` at variance 3, `cursor` at 3, `reve` at density2562 — all judged good. What separates a chosen low value from a timid one is that it is257executed to the end: fora commits to a single 1px rule and uses nothing else; reve gives one258capability the whole screen and never crowds it. A low value assumed and carried through259beats a high value claimed and not delivered, every time.260261**The middle is suspect.** Three 5s across three axes is the signature of a decision not262made. It is not banned — but it is worth knowing that **no page in the corpus carries a 5 on263more than one axis**. The two pages closest to the middle each buy it with a named device:264`umbrel` (5 · 6 · 4) with its solo/duet alternation, `mainframe` (5 · 4 · 6) with a brand265page that argues in the product register. So: if two or more dials land at 5, the plan says266in one line why the middle is the answer here, or moves one of them. Flag it as such — a267silent 5/5/5 is the failure this rule exists to catch.268269**A high value commits you to producing something.** The number is a promise the deliverable270has to keep:271272- `DESIGN_VARIANCE` ≥ 7 obliges a macrostructure that visibly departs from the canon **and**273 a signature procedure derived from the subject, applied more than once. `supercommon` is274 the only page here that clears 7, and it carries both: its solitary display line opens the275 page and then every band after it, then returns split in two as `stillness.` / `motion.`;276 and its voids are metered — each silence sized to the block behind it, the two largest277 (32vh, 40vh) reserved for the two blocks the page most wants read. Without both, high278 variance is novelty, not direction.279- `VISUAL_DENSITY` ≥ 7 obliges a counted inventory of what is on screen **and** two spacing280 regimes — dense inside blocks, airy between them. Every high-density page in the corpus281 splits this way. Density without the second regime is noise.282- `VISUAL_DENSITY` ≤ 3 obliges **the same counted inventory**, for the opposite reason: it283 has to show that something still holds each screen. See *The screen-holding floor* above,284 and the fourth column of the density table — both corpus pages at 2 are carrying real285 matter. A low density with nothing counted is not restraint, it is an empty page.286- `MOTION_INTENSITY` ≥ 6 obliges a documented motion system: the material list, the duration287 set, the curve set, and entrances differentiated by role rather than one blanket reveal.288 Above 7, add a named scroll mechanism and its `prefers-reduced-motion` path.289290If the deliverable does not carry the obligation, **lower the dial in the plan rather than291ship the number** — a claimed 8 that the page does not hold is the more expensive292mistake, because every downstream move inherits it.293294The macrostructure pick and `DESIGN_VARIANCE` must agree. A canonical skeleton is a 3; a295rebuilt-but-legible one a 5; dropped sections and a bespoke order a 7, which is where the296corpus stops — no page on disk shows what an 8 or a 9 costs to hold. A plan naming the297canonical skeleton next to `DESIGN_VARIANCE 8` has one of the two wrong — resolve it before298Gate 0, not while coding.299300## Gate 0 — Brief Lock301302Before routing to any target skill — before writing or modifying a single line of303HTML/CSS/tokens — six artefacts must exist **in writing**, not just in your head:3043050. **Register** — resolved above, stated explicitly (not defaulted).3061. **Tone** committed to ONE extreme — not an adjective that could also describe three307 competitors. Where it comes from depends on whether tone is one of the axes the308 Exploration Gate put in play:309310 - **Explored** — the gate fans out 3 divergent, text-only direction sketches judged311 blind and comparatively by `fuse-ai-pilot:challenger`, and the winner's tone becomes312 this artefact. The gate runs before Step 1 of `generate` **wherever that move has a313 page body to decide** — it is no longer a single scope/register special case.314 - **Not explored** — either no gate ran, or it ran on a reduced axis list because tone315 was already arbitrated in writing. Tone is then stated directly: `generate` derives it316 from the brief, refinement moves inherit it from the existing surface and name it.317 Inheriting is not skipping — an inherited tone left unstated fails this gate exactly318 like a missing one.319320 Which case applies, and which axes each scope × register combination puts in play, is321 defined **once**, in the scope table of `references/moves/generate.md` (§Exploration322 Gate). Read it there. **It is deliberately not restated here** — two copies of one rule323 are two rules that drift apart, and that file also owns the mechanics, the "reduced is324 not lighter" rule, and the fallback.325326 The fallback survives unchanged and is the only route back to an unexplored direction327 when the gate was due: `Task`/`Agent` unavailable at nesting depth 5 → the direct single328 pick, and the report marked "direction not explored / single-fiat". Never a silent one.3292. **Signature element** (brand register) or **primary task** (product register) named in330 one sentence.3313. **The corpus actually seen, then read** — in that order, and both halves are required:332333 - **at least two of the fifteen pages opened in a browser and scrolled through**, named in334 the plan as `{reference}-recode/index.html`. They open by double-click, `file://`, no335 server, no build;336 - `design-web/references/refs-design/README.md` plus at least two `tokens-*.md`337 sections, cited by reference and section name.338339 This is the evidence artefact, per `design-web/references/design-inspiration.md`.340 **Sections cited with no page opened does not satisfy it.** Measured values let you341 reproduce a procedure, never choose one, and a corpus of taste that is only ever read342 hands back conformity instead of judgment — the failure is documented at343 `design-web/references/refs-design/README.md` §*Look before you read*. A sector URL or a344 supplied screenshot is an addition to this artefact, never a substitute for either half.3453464. **What will be on screen, counted** — one list, block by block, of the images, videos,347 mockups, drawn interfaces, charts, tables, full-bleed colour fields and display-scale348 type objects the page will actually carry. Required unconditionally, and decisive349 whenever `VISUAL_DENSITY` ≤ 3 or `DESIGN_VARIANCE` ≥ 6: see *The screen-holding floor*350 above for what to do when the count comes up short. A plan that names a dial triple but351 cannot say what fills the screen has not been made.3525. **Structure named, both halves** — one line in the plan:353354 ```355 Macrostructure: {hero treatment from references/macrostructure-bank.md} + {body sequence from references/body-sequence-bank.md}356 ```357358 Naming one without the other fails this gate. The two banks are separate on purpose:359 the hero bank owns the **first screen only**, the body-sequence bank owns **everything360 after it**. The body, not the hero, is the documented failure mode361 (`references/register/brand.md` intro) — a plan that names a hero treatment and stops362 has improvised the scroll, and improvised scrolls converge on the same363 feature-grid-then-testimonials default the corpus was assembled to break. So an unnamed364 body sequence is a missing artefact, not a detail to settle while coding.365366 Restate the sequence's ordering principle in the brief's own terms, not the bank's367 label: "four chambers, one image transformation each" is a body sequence, "alternating"368 is a word. And the pick has to agree with `DESIGN_VARIANCE` (see the dial section369 above) — a canonical skeleton and a variance of 8 cannot both be true.370371 Where the Exploration Gate ran, this pair is **its output, not a fresh pick**: the372 winning sketch's hero treatment + body sequence, carried forward with its principle373 line. Recording it satisfies this artefact; re-deciding it discards the comparison that374 was just paid for. Where no gate ran, pick both from the banks here.375376This is a **present/absent check on six named artefacts** — structural, not a taste377judgment. One missing ⇒ **generation or modification is forbidden** until it exists.378Mechanics of artefacts 1–2 (the Exploration Gate and its scope table, brief questions,379signature element, two-pass critique) live in `references/moves/generate.md`; the two banks380own the vocabulary of artefact 5. This section is only the checkpoint every move respects381before it's allowed to touch code — it says *what must exist*, never *how it was produced*.382383## Design guidance384385Dense and general, applying whichever move is running — taste rules specific to one move386live in that move's file.387388### Absolute bans389390Match-and-refuse, unconditional across both registers — not a statistical reflex to weigh,391a hard stop. If you're about to write any of these, rewrite the element with a different392structure, never a variation on the same idea.393394- **Side-stripe borders** — `border-left`/`border-right` >1px as a colored accent on cards,395 list items, callouts, or alerts. Rewrite with full borders, background tints, leading396 numbers/icons, or nothing.397- **Gradient text** — `background-clip: text` combined with a gradient background.398 Decorative, never meaningful. Single solid color; emphasis via weight or size.399- **Glassmorphism as default** — blurs/glass cards used decoratively rather than rare and400 purposeful, or nothing.401- **The hero-metric template** — big number, small label, supporting stats, gradient402 accent. SaaS cliché; see `register/brand.md` §2 "Ribbon 4-stats" furniture for the fix.403- **Identical card grids** — same-sized cards, icon + heading + text, repeated endlessly.404- **Modal as first thought** — modals are usually laziness. Exhaust inline/progressive405 alternatives first.406- **Em dash in copy** — commas, colons, semicolons, periods, or parentheses instead. Also407 not `--`.408- **The AI-slop signature triad** — defined here and canonical here; every other file in409 the pipeline points back to this bullet instead of restating it. Three tells that each410 pass on their own and together fingerprint the page:411 1. a **Tailwind blue/indigo primary or accent**, hue range 200–290°;412 2. **Inter as the primary typeface**;413 3. **`rounded-2xl` as the default radius**.414415 That palette/font/radius combination is shared by **~83% of sampled AI-generated pages**416 (grounding: sailop ai-slop research). The figure belongs to this triad and to nothing417 else — never quote it for another ban, another detector, or a general anti-slop claim.418419 Any two legs are survivable; the third is what makes the page identifiable. So if brand420 guidelines already fix 2 of the 3, the third MUST change — another hue, another radius,421 or another typeface (reflex-reject list and alternatives:422 `design-system/references/forbidden-fonts.md`). The corpus carries the counter-example:423 `fora-recode` and `umbrel-recode` both ship Inter and both read specific, because neither424 carries either of the other two legs.425426Deterministic grep companions for the structural ones (side-stripe, identical grids,427nested cards, icon-bento): `design-review/references/anti-ai-slop-audit.md`.428429### The AI-slop test — 2 levels430431If someone could look at this interface and say "AI made that" without doubt, it's failed.432Run at **two altitudes** — the second catches what the first misses. Cross-register433failures are the Absolute bans above; register-specific failures live in each434`register/*.md`.435436**First-order — category reflex.** If someone could guess the theme + palette from the437category alone, it's the first training-data reflex:4384391. **Cream #F4F1EA + a contrasted serif + terracotta accent** — the default "editorial440 SaaS" look.4412. **Near-black background + one acid accent color** — the default "dark developer tool"442 look.4433. **Broadsheet hairlines, zero border-radius, black/white only** — the default "premium444 minimal" look.4454. **Glassmorphism + `rounded-2xl` used globally** — the default "2026 AI app" look,446 applied everywhere instead of gated (`design-motion` gates it deliberately).4475. **Generic icon-bento** — every cell centered text over a round colored-icon badge, zero448 image/gradient/pattern variation between cells.449450Rework the Design Read's vibe words and the color strategy (`design-system` §Color451strategy) until the answer isn't obvious from the category — "observability → dark blue",452"healthcare → white + teal", "finance → navy + gold", "crypto → neon on black" are all this453tier. Purple-on-white gradients are banned outright regardless of tier — the single most454common tell.455456**Second-order — category-plus-anti-reference trap.** The first reflex was avoided, but if457someone could still guess the aesthetic *family* from category-plus-anti-references — "AI458workflow tool that's not SaaS-cream → editorial-typographic", "fintech that's not459navy-and-gold → terminal-native dark mode" — it's the trap one tier deeper. Rework until460neither answer is obvious. Register-specific reflex-reject aesthetic lanes:461`register/brand.md` and `register/product.md`.462463Deterministic grep detectors cover the first-order clusters plus the Absolute bans; the464second-order trap is judgment, checked at the challenger gate465(`design-review/references/review-procedure.md`), not grep-lintable.466467### Macrostructure variety — hero AND body468469Centered hero + 3-column icon-card grid is forbidden as a default skeleton. Before any plan470is finalized, name **two** picks: a hero treatment from `references/macrostructure-bank.md`471(first screen only) and a body sequence from `references/body-sequence-bank.md` (everything472after it, with its ordering principle restated in the brief's own terms). A plan carrying473only a hero treatment has improvised the scroll — the body is the documented failure mode.474475Following the canon is a legitimate pick. `fora-recode`'s "Bookended Canon" runs the476canonical skeleton almost to the letter and does not read generic, because one device — a477single 1px rule at white 10% — is carried through every edge, card and divider.478`endlesstools-recode` keeps the same order and fills every slot with footage. Departing is479equally legitimate: `supercommon-recode`'s "Void-Metered Stack" drops the testimonials, the480pricing table, the FAQ, the CTA band and the nav outright — the price is inline on one481button and the exit is a single link with no headline above it. What is never legitimate is482**not choosing** — the skeleton that arrived by default is the one that reads as generated.483484Let each pick follow the subject, not habit: a treatment or a sequence reused by reflex485across briefs becomes its own template, whichever one it is. A pick used on two consecutive486briefs has to be argued from the second subject or replaced — the fact that it worked once487is not an argument.488489### Non-negotiable floor490491Regardless of tone, register, target, or move, every deliverable must have: responsive492behavior across the target's breakpoints/size classes; visible keyboard focus on every493interactive element (`:focus-visible`, never suppressed); `prefers-reduced-motion`494respected wherever motion is added.495496### Generation approach497498Generate HTML/CSS directly — the default and primary path, following the same method as499Anthropic's official `frontend-design` skill (commit to a point of view, avoid templated500defaults, verify with tools not vibes). Gemini501502…(truncated)