Redesign an existing website to make it better. Built on top of impeccable.
88%
Does it follow best practices?
Run evals on this skill
Adds up to 20 points to the overall score
Low
Low-risk findings worth noting
The text of this part of skills/deploy/SKILL.md, which keeps the heading and points here. "Above" and "below" refer to the order of the sections in that file.
Dispatch one agent per page-archetype cluster (utility pages, services, case studies, etc.). Each agent owns a non-overlapping set of new blocks and content pages. Three to four parallel agents is the sweet spot.
Brief size and reading discipline. The brief points at files — stardust/eds-schema/<page>.json, the conversion log's triage rows for its pages, this document's §§ 7–8, davids-model.md — and never pastes reference text into the prompt. Each agent reads by section (list the headings, then read the range it needs), not the whole file: across twelve field migrations this document (~27k words) was read end to end about twenty times per run, once per dispatched agent; in one recorded run the two conversion agents the harness's no-progress watchdog killed carried the fattest briefs, while a re-dispatch with a lean brief and line-ranged reads finished the same pages. Long-running steps (captures, gates, batch pushes) run in the background with a progress file the agent appends to per page, so the coordinator can read progress instead of waiting blind. The coordinator's own waiting follows the master skill's wait discipline: nothing runs in the foreground past ~2 minutes, no single sleep reaches 5 minutes (the prompt-cache window — at deploy-phase context sizes each expiry re-writes the whole prefix), and progress is read from the file at most every 4 minutes.
The brief template:
Per the project's locked direction: each prototype
<section>becomes its own EDS block. Lift the prototype's<style>for that section verbatim, scope it under the block class (.block-name .xinstead ofsection.x .y), and rebuild the prototype's DOM through adecorate(block)function that consumes EDS table-block input.You own: prototypes [list], content pages [list], sections [list].
Existing blocks — REUSE, do not recreate: [list with one-line authoring shape per block].
Brand tokens are global in
styles/styles.css; do not redefine.Round-trip contract (#93/#94): the page's authored rows AND your block's decode are both written from
stardust/eds-schema/<page>.json(roles + repeat units — Step 2b); cite the schema path in the block JSDoc. After writing each block, runnode skills/deploy/scripts/block-roundtrip.mjs "<protoURL>" content/<page>.html --blocks <name>— the block is NOT done until it exits 0 (0 structural 🔴).Section layout — reproduce the prototype's max-width container (#13): if the prototype section wraps its content in a centered max-width container (
<div class="wrap">/.container/.inner), your block MUST recreate it — build the content into a.wrapdiv (block.replaceChildren(wrap)), so the colored/section background bleeds full-width but the content stays within the page max-width. Only render content edge-to-edge where the prototype section itself is full-bleed (no inner wrapper). Getting this wrong is invisible at ≤1440px and only shows at wide viewports.Breakpoint overrides must match variant specificity (#109): write every breakpoint override at the variant's own specificity (one override per variant, or
:where()the variant selectors down). The trap is silent on single-variant blocks and bites the moment a rollout adds variants. One more trap: a wrapper reset can out-specify the block's own rules (#114) —footer .footer > div { padding: 0 }beatsfooter .f-root { padding: … }silently; keep resets at LOWER specificity than the rules they might shadow (:where()them down).Flex/grid children — no
<br>, no bare child rules: inside a flex or grid container an authored<br>becomes a SIZED flex item (recorded: a two-line label splayed across a 120px tile by a ~26px<br>item) — emit separate elements instead. And never write a bare.label > span { display: block }-style child rule in variant CSS: it out-specifies a lower-specificity.explore { display: none }and resurrects hidden elements — scope with:not()against every state class the block owns, and verify with a computed-style dump of the child list, not by eye.Images —
<image-slot>placeholders (#2): claude-design prototypes use<image-slot>custom elements as image drop-targets; there are usually NO real image assets. Treat each image as an optional authored cell holding a<picture>/<img>(const pic = cell.querySelector('picture, img'); if (pic) …). When the cell is empty, fall back to the prototype's background treatment (e.g. dark--ink, or a placeholder rectangle) via the block CSS so the section still looks right with no image. Leave image cells EMPTY in the authoring snippet.Scroll-reveal / JS-hidden content (#14): if the prototype hides content behind a class an inline
<script>toggles on scroll (.reveal { opacity:0 }+ an IntersectionObserver that adds.in), do NOT lift theopacity:0— the prototype script does not run in EDS, so the content would be permanently invisible. Render it visible; drop the reveal (keep only hover/:hovertransitions). Honorprefers-reduced-motion.Interactive / component-driven sections (#17): when a section is driven by a component (state, a list loop like
<sc-for>, conditionals like<sc-if>,{{ }}bindings, adata-countcounter, a tab/selector), split it: data → authorable rows (one row per list item, with the item's fields as cells) and behavior → block JS. Unlike static fragments, block JS runs — sodecorate()is the right place to wire click handlers, an IntersectionObserver count-up, tab switching, etc. Render the default/active state in markup; drive the rest from JS-held local state.{{ }}/<sc-for>/<sc-if>are NOT EDS syntax — read them as "loop these rows" / "show one state". The same split applies to blocks whose data arrives at runtime —stardust/dynamic-features.mdrows with dispositiondata-fedorindex-backed: the block's authored rows ARE the fallback (rendered before any fetch resolves, and what ships if it never does),decorate()fetches theSource/ index and re-renders, and the brief names the endpoint, the fields the cards need, and the failure behaviour. Never author a block whose only content arrives by fetch — and when NO authored rows remain (search, a filter, a calculator, a mount: the UI itself, not content) the delivery form is a widget (Step 2), not a block; widgets are scoped to non-content UI, so this document-first rule still governs every listing. The inventory is brief input for every page:#modallink markers, player URLs, per-page form endpoints inscripts/site-config.js. Patterns:skills/dynamics/reference/patterns.md.David's Model (the authored-structure contract —
davids-model.md): a prose section with no repeating units and no bespoke structure is DEFAULT CONTENT, not a block (D1 — the Step-2 triage in the conversion log says which of your sections these are); no nested block tables (D2); blocks stay ≤4 columns (D10); if your section matches a Block Collection pattern the conversion log names, follow that block's authoring shape (D11); no code visible as text in cells (D15). Your pages must passnode skills/deploy/scripts/davids-model-lint.mjs content/<page>.htmlwith 0 🔴.Buttons: do NOT manufacture button anchors. Author CTAs paragraph-wrapped as
<strong><a>(primary) or<em><a>(secondary) in the content page —decorateButtons()classes them (a.button.primary/.secondary) BEFORE yourdecorate()runs; in block JS, MOVE the CTA paragraph into a.actionswrapper (actions.append(a.closest('p') || a)) — never clone it (EW3). Block CSS only overrides global button styles when something is genuinely different (e.g. larger size). Text links with flourish (wavelength underline) are NOT buttons — leave as plain<a>and style per-block.Experience Workspace editability contract (Step 8, EW1–EW10) — every block you write obeys it and passes the EW gate before it is done. MOVE authored
h1–h6/p/ul/ol/pictureinto generated wrappers (wrap.append(heading)); never rebuild fromtextContent/innerHTML, nevercloneNode+ discard, never retag, never.trim()displayed text (EW1). Wrappers carry the layout classes; style the authored element through the wrapper with DESCENDANT selectors (.headline :is(h2, h3)), never a class on the authored element and never>/:first-child/:nth-childon the path to it (EW2). CTAs move as their<p>(EW3). Presentational clones (loop slides, marquee copies) getstripInstrumentation()(EW4). Text you cannot make editable is declared with an@ew-exemptJSDoc tag, never dropped silently (EW5). A<button>/<summary>cannot host the editor (EW7).block-roundtripruns with--ewby default: a dead non-exempt text or a duplicated index is a 🔴.EDS block convention: each block at
blocks/<name>/<name>.{js,css}; each widget atwidgets/<path>/<name>.{html,css,js}(+ optional.json), loaded by the stockblocks/widget/. JS exportsdefault async function decorate(block). Block input is<div class="block-name"><div>row<div>cell</div></div>…</div>(the runtime adds.block+data-block-nameand nests it in.<name>-wrapperbefore your JS runs). CSS scoped under.block-name. Inline SVG markup per-block (no shared utility). Honorprefers-reduced-motion.EDS content page format: NO
<head>element (projecthead.htmlis injected by EDS), empty<header></header>/<footer></footer>, each top-level<div>inside<main>is one section holding one block OR default content, section-metadata only as a Step-3stylevalue on default-content sections, no<style>/<script>, fully-qualified image URLs.Done criteria: [list of paths]. Return a list of new blocks + one-line summary per page.
Agents do not need to coordinate on shared blocks — the brief tells them which existing blocks to reuse.
.tessl-plugin
skills
audit
reference
deploy
reference
scripts
test
diff
direct
dynamics
extract
migrate
prepare-migration
prototype
assets
motion
fixtures
reference
qa
replica
reference
scripts
test
reskin
rollout
spec
stardust
uplift
reference