stardust:distill

Read a directory of user-provided design samples — static HTML files the user has placed under samples/static/, plus a samples/live/urls.json manifest of live URLs to fetch — and produce two artifacts that anchor the rest of the stardust pipeline:

1. samples/trait-matrix.json — structured per-sample + cross-sample trait data (type system, color, spacing, motion, component patterns, voice, assets).
2. samples/SAMPLES.md — a narrative vocabulary brief, the file stardust:direct reads under Mode B as the resolved anchor set.

This skill is descriptive and pre-extract. It does not crawl an origin (that is extract's job). It does not invent design opinions (that is direct's job). It produces the brief the user hands to direct as Mode B anchor references when they want stardust to follow a specific visual language — captured from samples — instead of rolling the divergence seed cold.

Distill is optional. Users who do not have samples skip straight to extract. Distill becomes load-bearing when the user has been instructed by a brand team, a designer, or an art director to honor specific reference material that is not the existing site (the common case: a refresh whose visual direction has been pre-set by design leadership and arrives as a folder of HTML + screenshots).

Inputs

• <samples-dir> — optional positional. Defaults to samples/.
• --from-static — distill only static samples; skip live fetching. Use when no live URLs are listed or when the user wants to defer the live pass.
• --from-live — distill only live samples; skip static parsing. Use when the static pass has already run and only live snapshots need refreshing.
• --cap <N> — cap live URLs to fetch (default 8). Matches extract's small-sample philosophy: 8 reference URLs is enough to triangulate a visual language without burning budget on near-duplicates.
• --refresh — re-fetch and re-parse every sample even if cached snapshots exist under samples/live/snapshots/. Default behaviour is to reuse cached snapshots and only fetch URLs that are missing.

Setup

Run the master skill's setup procedure first (skills/stardust/SKILL.md § Setup): impeccable dep check, context loader, state read.

Additional checks for this sub-command:

1. Playwright availability — same dep check as extract. The live pass needs a real browser. Detect Playwright in this order: a Playwright MCP server, then npx playwright. If neither is available and --from-static was not passed, stop and tell the user how to install Playwright.
2. Samples directory structure — verify the input dir exists and contains at least one of: static/<sample>/, live/urls.json. If neither, stop and explain the expected layout:

   samples/
     static/
       <sample-1>/
         index.html         (or any .html file)
         assets/            (optional; reused if present)
       <sample-2>/
         ...
     live/
       urls.json            (array of { url, slug?, role? } objects)
   

3. State precedence — if stardust/state.json already records a samples.distilledAt timestamp and --refresh was not passed, ask whether the user wants to re-distill (overwriting) or skip. Distill is idempotent; re-running it overwrites the prior brief and trait matrix.

Procedure

Phase 1 — Inventory the static samples

For each subdirectory of samples/static/:

1. Identify the entry HTML file (index.html if present, otherwise the largest .html file by size).
2. Read the file and its referenced stylesheets (inline <style> + linked <link rel="stylesheet"> resolved relative to the sample dir, both samples/static/<sample>/... and any sibling assets/ directory).
3. If a screenshot or PDF lives in the sample dir, note its path — reviewers may want to compare against the parsed traits.
4. Record an entry in the in-memory inventory: { slug, kind: "static", entryHtml, stylesheets[], screenshot?, role? }. The role (if any) is inferred from the directory name (bizpro-hub-prototype → business; plan-page → pricing; etc.) — store the inference, surface to user for correction.

Phase 2 — Fetch the live samples

If live/urls.json exists and --from-static was not passed:

1. Read the manifest. Each entry: { url, slug?, role? }. Auto- generate slug from the URL hostname + path if missing.
2. For each URL (capped per --cap):
   • Check cache: if live/snapshots/<slug>.html + .png exist and --refresh was not passed, skip the fetch and use the cached pair.
   • Otherwise, render with Playwright following extract's playwright-recipe.md: viewport 1440 × 900 @ 2× DPR, wait mode medium, reduced motion, scroll-to-bottom pass, capture the full DOM + a full-page screenshot.
   • Save:
     • live/snapshots/<slug>.html (rendered DOM)
     • live/snapshots/<slug>.png (full-page screenshot)
     • live/snapshots/<slug>.log (one-line provenance: { url, fetchedAt, waitMs, status })
3. Record an entry in the inventory: { slug, kind: "live", url, snapshotHtml, snapshotPng, snapshotLog, role? }.

Phase 3 — Per-sample trait extraction

For each entry in the inventory, parse the HTML + computed styles (use a headless browser pass with getComputedStyle() on representative elements — same technique as extract's per-section style summary). Extract per-sample:

• Type system — every named CSS variable matching --*font-*, --*type-*, plus computed style on every heading, body paragraph, button, eyebrow, link. Aggregate into { headingFamily, bodyFamily, labelFamily, weights[], scale: { title-1, title-2, ..., body-lg, body-sm, eyebrow, label }, rules[] }. Detect modular scale per extract/reference/brand-surface.md § Modular-scale audit.
• Color — every named CSS variable matching --*color-*, --*bg-*, --*surface-*, --*text-*, plus computed background- color and color on representative elements. Aggregate into { palette: { neutrals[], brand: {}, surfaces: {} }, reservations: [{ color, reservedFor[] }], rules[] }. The reservedFor list is inferred from class names + CSS selectors (e.g., .brand-mnemonic ⇒ reservedFor: ["brand-mnemonic"]).
• Spacing — every named CSS variable matching --*space-*, --*spacing-*, --*pad-*, --*gap-*. Detect density tier (compact / balanced / packed) by clustering section padding values per intent-dimensions.md § 4.
• Radius / shadow / motion tokens — every named CSS variable matching --*radius-*, --*shadow-*, --*ease-*, --*duration-*, --*motion-*.
• Motion stack — detect loaded libraries via <script src=> inspection: GSAP, ScrollTrigger, Lenis, Locomotive, Three.js, Splide, Swiper. Pin versions where readable from the URL. Detect CSS-only motion patterns by scanning for animation-timeline:, @keyframes, transition:. Record:

  {
    libraries: [{ name, version?, role }],
    cssOnlyPatterns: ["animation-timeline: view(block)", ...],
    choreographies: [{ name, surface, mechanism, capturedFrom }]
  }
  

• Component patterns — detect repeating DOM signatures: card grid, banner, navigation anatomy, hero composition. Per extract/reference/brand-surface.md § System components. For each pattern, record { kind, shape, instanceCount, examples[] } — the examples include enough markup snippet for direct to understand the captured shape.
• Voice samples — the hero headline, first body paragraph, three representative CTAs, three navigation labels, the footer line. Same fields as extract's voice block; the difference is voice is per-sample here (not aggregated cross-sample, since samples may represent different registers).
• Asset inventory — fonts (@font-face declarations + their src URLs), images (with intrinsic dimensions), inline SVGs. Per-sample; cached locally if accessible.

Save the per-sample data into the trait matrix's samples[] array.

Phase 4 — Cross-sample aggregation

Walk the per-sample data and produce the sharedTokens block. A token is "shared" if at least ceil(N/2) samples carry the same value (after near-duplicate clustering for colors). Otherwise it goes into divergentTraits with a per-sample breakdown.

Aggregate:

• sharedTokens.type — type families, weights, scale that recur.
• sharedTokens.color — neutrals + brand + surface tokens that recur; reservations are union'd (one sample's reservation is enough to honor).
• sharedTokens.spacing — section padding and grid that recur.
• sharedTokens.radius / shadow / easing / duration — modal values.
• sharedTokens.motionStack — libraries loaded by ≥1 sample (motion is opt-in, not majority-rule).
• divergentTraits — anything that didn't reach the majority threshold. Direct reads this to know which traits to amplify in the B/C/... variants (the "captured trait amplified" role).

Phase 5 — Author SAMPLES.md (narrative brief)

SAMPLES.md is the load-bearing artifact direct reads under Mode B. Format per reference/samples-brief-format.md:

• Provenance header (HTML comment as first line) — same shape as every other stardust artifact provenance block: writtenBy, writtenAt, readArtifacts, samples, stardustRole.
• Opening paragraph — one paragraph framing the sample set as a reference family. Names the shared design language, the per-sample registers, and the consumer of this brief (direct, Phase 2 Mode B).
• § 1 Type system — table of the shared scale; rules block; fluid clamp() formulas if detected.
• § 2 Color — palette as a swatched table; reservations list; brand-faithful inversion candidates (per direct/reference/direction-format.md § Brand-faithful inversions).
• § 3 Spacing + grid + density tier.
• § 4 Motion stack + choreographies — library list with versions; named choreographies (hero scroll-grow, garage-door, etc.) with the mechanism + sample they were captured from.
• § 5 Component patterns — cards, banners, navigation, hero composition; one shape brief per pattern.
• § 6 Voice per register — sample-by-sample voice block (not aggregated); direct uses the per-sample voice to match register in B/C variants.
• § 7 Asset inventory — fonts (with license flags), photography vs illustration ratio, signature motifs.
• § 8 Divergent traits — the list direct uses to pick which trait to amplify in B/C variants.

Author directly — distillation does not need impeccable's authoring ceremony. SAMPLES.md is consumed by direct as factual input, not as design opinion.

Phase 6 — Update state and report

After all phases succeed:

1. Update stardust/state.json (add a samples block if absent):

   "samples": {
     "distilledAt": "<ISO-8601>",
     "samplesDir": "samples/",
     "brief": "samples/SAMPLES.md",
     "traitMatrix": "samples/trait-matrix.json",
     "inventory": [
       { "slug": "bizpro-hub", "kind": "static", "role": "business" },
       { "slug": "sr-homepage", "kind": "live", "url": "...", "role": "consumer-creative" },
       ...
     ]
   }
   

2. Print a one-screen summary:

   Distilled 3 samples (2 static, 1 live)
   
   samples/
     SAMPLES.md             (8 sections, 4 motion choreographies named)
     trait-matrix.json      (12 shared tokens, 4 divergent traits)
     live/snapshots/        (1 file)
   
   Per-sample evidence:
     slug          kind     role                 typeFam               motion-libs
     bizpro-hub    static   business             adobe-clean-display   Lenis 1.x, GSAP 3
     plan-page     static   pricing              adobe-clean-display   Lenis 1.x
     sr-homepage   live     consumer-creative    adobe-clean-display   Lenis 1.x, GSAP 3
   
   Shared tokens: type (12), color (8), spacing (10), motion (2 libs)
   Divergent traits (amplification candidates):
     - scroll-driven choreography (bizpro-hub strongest)
     - photography-led tile vocabulary (sr-homepage strongest)
     - pricing-tier card density (plan-page strongest)
   
   Next: $stardust direct "<phrase referencing SAMPLES.md>" --variants 3
   

When distill runs vs when it does not

Distill is optional. Three common scenarios:

1. No samples, just a refresh. User runs $stardust extract → $stardust direct directly. Mode A (brand-faithful) handles the common case — palette and type are pinned to the captured site, variants explore expressive / density / motion axes within the captured surface. Distill is not invoked.
2. Samples present + brand-faithful refresh. User runs distill first, then extract, then direct "<phrase referencing SAMPLES.md>". Mode B's anchor-reference precedence (per direct/SKILL.md § Mode B) reads SAMPLES.md as the resolved anchor set; Mode A's palette + type inheritance still applies because the captured brand wins on those axes. Distillation contributes the non-brand axes: motion stack, component shapes, density tier, choreography names.
3. Rebrand from a sample set. User runs distill + direct --rebrand "<phrase>". Distillation supplies the entire palette
   • type + motion language; extract's captured surface is only used for IA + content. This is the rarest case but the one distill was originally written for.

Inputs the user prepares (out of scope for this skill)

Distill consumes a directory the user assembles by hand. Common sources for samples:

• A static HTML mock from a designer (open in browser, save full page).
• A live URL provided by the brand team — list these in samples/live/urls.json; distill fetches them.
• A printed PDF / Figma export converted to HTML (best-effort).

Distill does not curate samples for the user; it does not invent sample sources; it does not fetch arbitrary URLs the user didn't list. If the samples are wrong, the brief will be wrong; this is intentional — distill is a faithful descriptor, not a curator.

Provenance

Every artifact distill writes carries a provenance block with at minimum: writtenBy: stardust:distill, writtenAt: <ISO-8601>, readArtifacts: [...] (every sample file consumed), consumedBy: "stardust:direct — Mode B anchor resolution".

References

• reference/samples-brief-format.md — the narrative-brief format the SAMPLES.md output follows.
• reference/trait-matrix-schema.md — the per-sample and cross-sample schema the trait-matrix.json output follows.