Skip to content

Web — Marginalia, Pandoc, Static Capture, Data-URI

The editorial-HTML channel and its tooling. Covers:

  • Marginalia — the typographic callout library for blog posts and paper-style pages
  • Pandoc bridge — the Lua filter that emits marginalia classes from markdown
  • Web rendering & static capture — Playwright / weasyprint / headless Chrome for DOM → PNG / PDF
  • Data-URI embedding — single-file portability trick for artifacts that must travel self-contained

Part of the muriel skill — see the top-level index for mission, universal rules, and channel map.

Marginalia — Editorial HTML

The editorial layer for blog posts, paper-style web pages, and any prose that needs typographic callouts. 15 components, zero dependencies, dark theme by default. CSS-only — JS is optional.

Library: marginalia on npm (CDN: marginalia@latest via jsDelivr, see below). Source and full pattern catalog in the andyed/marginalia repo; the repo's SKILL.md and llm.md are the reference docs when depth is needed.

Quick setup

<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/marginalia@latest/marginalia.css">
<script src="https://cdn.jsdelivr.net/npm/marginalia@latest/marginalia.js" defer></script>
<script src="https://cdn.jsdelivr.net/npm/marginalia@latest/marginalia-md.js" defer></script>

Markdown patterns (subset)

> [!NOTE]    > [!TIP]    > [!WARNING]    > [!IMPORTANT]
> [!QUOTE]   > [!ASIDE]  > [!MARGIN]
==highlighted==    {Badge}    {Badge: tip}    [^1](footnote)
{dropcap}

Plain > blockquotes become 3D perspective pull quotes — the signature component. Wrap content in <div class="mg-spread"> for two-column magazine layout.

Rules (from past feedback)

  • Core narrative inline always. Sidebars are for "go deeper" references only — one per section, at the heading.
  • Decorative elements ≥3:1 contrast. Don't fade below — compute the ratio.
  • All text 8:1 minimum.
  • All marginalia classes use mg- prefix. Theme via --mg-* CSS custom properties — these inherit into child SVGs, so a single theme switch repaints the whole page including embedded vector graphics.

Math — KaTeX

Marginalia essays with math drop KaTeX in via CDN + auto-render. MIT, no bundler, no MathJax-style runtime overhead. Pin to ^0.16. Display math is $$…$$, inline is $…$. Wrap display equations in a .eq-block container themed from --mg-* variables so equations participate in the marginalia palette.

Reference exemplar: inside_the_math — math-heavy essay, marginalia + KaTeX + WebGL demos. Full vocabulary in vocabularies/katex.md: setup, the .eq-block pattern, color/emphasis with \textcolor, server-side rendering for stills.

Do not use KaTeX for paper figures — see channels/science.md, where matplotlib + usetex=True keeps math inside the same TeX pass as the rest of the figure. KaTeX is the web answer; LaTeX is the paper answer.

Pandoc → Marginalia Bridge

Pandoc 3.2+ recognizes GitHub-style alerts (> [!NOTE]) via the +alerts extension. A Lua filter at <marginalia>/pandoc/marginalia.lua rewrites pandoc's output to marginalia classes: GitHub alerts → mg-callout[data-type], extended markers (ASIDE/MARGIN/QUOTE) → mg-sidebar / mg-margin / mg-pull, plain blockquotes → mg-pull, ==mark== → mg-mark, {Badge} and multi-word {Text: type} → mg-badge, pandoc footnotes → mg-fn popover, fenced code → mg-code, inline code → mg-inline-code, {dropcap} line marker → mg-dropcap wrap.

Usage

pandoc draft.md \
  --from markdown+mark+alerts \
  --lua-filter <marginalia>/pandoc/marginalia.lua \
  --template <marginalia>/pandoc/template.html \
  --standalone \
  -o draft.html

Same source fans out to PDF (via weasyprint), DOCX, or EPUB without losing marginalia's typographic voice:

# HTML → PDF
weasyprint draft.html draft.pdf

# Direct to DOCX (filter still applies)
pandoc draft.md --from markdown+mark+alerts \
  --lua-filter <marginalia>/pandoc/marginalia.lua \
  -o draft.docx

Reference example

The us-constitution.md example in the marginalia repo — an annotated US Constitution reader's edition. Exercises every transformation: dropcap, all four callout colors, sidebars, margin notes in the desktop gutter via CSS Grid subgrid, pull quote, inline highlights, badges, footnote popover, fenced code, plus a cursive copperplate dropcap initial and closing signature paragraph. Good starting point for any new marginalia page.

Light editorial variant — the F explainer pattern

The Attentional Foraging F-pattern explainer is the reference exemplar for a warm, reader-friendly light mode that extends marginalia beyond its default OLED palette. Reach for this pattern when a project wants explainer pages, blog posts, or paper drafts on a warm cream background instead of dark.

Note on status: the explainer is live and in active use, but some supplementary sections are still in draft; treat specific claims as provisional until the corresponding draft is lifted.

Build path

The F explainer uses the browser-side marginalia-md.js converter via a Node build script, not pandoc. Different path from the pandoc bridge above, same underlying marginalia classes — the build script also post-processes {dropcap} into <p class="has-dropcap"> as a workaround for a marginalia-md rendering quirk.

When to use each path: - Pandoc bridge — for paper-draft round-trips (markdown → HTML → PDF → DOCX), when the same source needs to fan out to multiple output formats - marginalia-md.js + Node build — for single-target static explainer pages with heavy custom extensions and inline raw HTML

Extensions beyond core marginalia

The F explainer adds four non-canonical patterns. They're worth knowing by name because they solve gaps marginalia doesn't cover:

Class / pattern Purpose Why it's not in core marginalia
.outer-note Short punchy margin labels (0.75em, amber #b08050, ~150px wide, position: absolute; right: -180px). Hidden at max-width: 1100px. mg-margin is for full-sentence annotations; these are one-phrase rubrications that work like Tufte sidenotes.
.stats-detail Inline span for numerical details with a dashed amber underline (ρ = 0.02, p = .30). Warm cream background, smaller font. No native inline-stats treatment in marginalia; inline <code> is wrong semantically.
.has-dropcap on <p> directly {dropcap} line marker post-processed to <p class="has-dropcap">. CSS targets p.has-dropcap::first-letter. Works around a marginalia-md.js rendering quirk; conceptually equivalent to our pandoc filter's mg-dropcap div wrap.
Staged h2 colors via :nth-of-type h2:nth-of-type(1..4) get gold/red/blue/green for the OSEC stages (Orient, Survey, Evaluate, Commit). Document-structure-driven color. Works because this page's h2s are the OSEC stages in order, so position is semantics.

The F explainer also floats .mg-margin and .mg-sidebar to the right inside the content area (230px width, amber-tinted #d4a574 border) rather than escaping to the outer gutter the way the US Constitution example does. Reader-friendly on narrower viewports.

Warm editorial palette

Token Value Role
Background #fafaf8 Warm cream, not pure white
Body text #222222 Near-black on Georgia serif
Accent amber #d4a574 Margin-note borders, legend edges
Note blue #f5f8fd bg + #2266cc80 border Info callouts
Tip green #f2faf4 bg + #22883380 border Positive notes
Warning amber #fdf9f0 bg + #cc880080 border Cautions
Important red #fdf5f5 bg + #cc444480 border Emphasis
Stats-detail bg #f8f4ee + dashed #d4a574 underline Inline numerical callouts

Matching matplotlibrc: muriel/matplotlibrc_light.py. Same Georgia stack, same #fafaf8 figure facecolor — figures rendered with the light rcparams match the editorial palette 1:1.

When to use which variant

  • Dark (OLED) — default. Blog posts on dark-themed sites, paper figures destined for a dark slide deck, any page using the default data-mg-theme="dark".
  • Light editorial — long-form explainers, paper drafts in light review UI, blog posts on light-themed sites. Match to the F explainer when the voice is "reader's edition" not "research console".

Pick one per document. Don't mix dark and light callouts in the same paper or post.

Web Rendering & Static Capture

The "DOM is the rendering engine" trick: build the artifact as HTML/CSS/SVG/JS, then capture it with headless Chrome or Playwright. Combines marginalia's editorial chrome, SVG's vector precision, and interactive JS's parameter sliders — then freezes one state into PNG or PDF for distribution.

When to use

  • Capture an SVG that's CSS-themed at runtime (avoids redoing colors per export)
  • Snapshot an interactive demo at a specific parameter state
  • Render a marginalia paper-style HTML page as PDF for paper submission
  • Build infographics in HTML (CSS Grid, marginalia callouts, embedded SVG) and rasterize for store/social
  • Capture a Three.js scene at exact framing for blog hero images
  • Generate device-frame screenshots from real page renders rather than mocking in Photoshop

Tooling

Tool Strength Recipe
Playwright (Python or Node) Full automation, waits for fonts/network/JS, supports clip regions page.goto(url); page.locator('svg').screenshot()
Puppeteer Same idea, Chrome-only, lighter await page.screenshot({ path, clip })
chrome --headless --screenshot Zero-script one-off chrome --headless --screenshot=out.png --window-size=1280,720 file://...
weasyprint HTML+CSS → PDF without a browser; respects @page rules weasyprint input.html out.pdf

Patterns

  • Wait for fonts. page.evaluate(() => document.fonts.ready) before screenshot, or webfonts render mid-capture.
  • Set device_scale_factor=2 for retina-quality output.
  • Use clip regions to capture a single SVG/canvas instead of the whole viewport.
  • Inject parameters via URL hash so the same demo HTML can be captured at multiple states (pairs with PermalinkManager pattern).
  • Use --virtual-time-budget in headless Chrome to fast-forward animations to a stable frame.
  • prefers-color-scheme: dark matters. Force it via page.emulate_media(color_scheme='dark') so OLED palette renders correctly.
  • Capture small multiples in one run. Loop over parameter values, set URL hash, screenshot. Output a grid of frames that compose into a small-multiples figure.

Example: SVG → PNG via Playwright

from playwright.sync_api import sync_playwright
with sync_playwright() as p:
    browser = p.chromium.launch()
    page = browser.new_page(device_scale_factor=2)
    page.set_content(open('chart.html').read())
    page.evaluate('document.fonts.ready')
    page.locator('svg#chart').screenshot(path='chart.png')
    browser.close()

Responsive sweep helper — muriel.capture

For "what does this look like at mobile/tablet/laptop/desktop?" in one command, the paired module muriel/capture.py wraps Playwright around the tier data in muriel/dimensions.py:

from muriel.capture import capture_responsive

paths = capture_responsive(
    url="https://example.com/",
    output_dir="captures/",
    color_scheme="dark",
)
# Writes four retina PNGs (device_scale_factor=2):
#   captures/example-com-mobile-390x844.png
#   captures/example-com-tablet-820x1180.png
#   captures/example-com-laptop-1280x800.png
#   captures/example-com-desktop-1920x1080.png

Or from the command line:

python -m muriel.capture https://example.com/
python -m muriel.capture https://example.com --dir captures/ --dark
python -m muriel.capture https://example.com --selector main --slug hero
python -m muriel.capture https://example.com --tiers mobile-large desktop-qhd ultrawide
python -m muriel.capture https://example.com --full-page

Defaults: four-tier sweep (mobile tablet laptop desktop), device_scale_factor=2, wait for document.fonts.ready, wait for networkidle. Output filenames follow the naming convention in channels/dimensions.md. Exit 0 on full success, 1 if any tier failed, 2 on missing Playwright / Chromium or usage error. Playwright is an optional install — the module is importable without it and raises a clean ImportError with install instructions on first call to capture_responsive().

Example: marginalia HTML → PDF via weasyprint

weasyprint paper-draft.html paper-draft.pdf
Add @page { size: A4; margin: 2cm } to the marginalia stylesheet override for paper formatting.

Static capture as the unifier

This channel is what makes the whole toolkit composable. SVG rendered with CSS theming → captured as PNG → embedded in a video. Interactive demo at three parameter states → captured as small multiples → assembled into one image. Marginalia page → captured as PDF for submission. The DOM is the most flexible compositor; static capture is the export step.

Data-URI Embedding

The portability trick: embed images directly inside HTML/SVG/CSS as data:image/png;base64,... URIs. Single-file artifacts with zero asset management — they email, paste into Slack, copy onto USB sticks, embed in iframes from sandboxes that block external loads.

When to use

  • Single-file HTML demos (one file = the whole thing)
  • SVG with embedded raster textures (one file = the whole figure)
  • Marginalia blog post fragments that need to be paste-able
  • Email-friendly artifacts (Outlook/Gmail strip external <img> until clicked)
  • Paper figures where the journal wants a self-contained submission
  • CSS backgrounds without HTTP round-trips
  • Inlining icons inside SVG so a single file renders correctly when downloaded

Recipes

Bash → data URI: 

# PNG → data URI string
echo "data:image/png;base64,$(base64 -i input.png | tr -d '\n')"

# SVG → data URI (URL-encoded, smaller than base64 for SVG)
python3 -c "import urllib.parse; print('data:image/svg+xml,' + urllib.parse.quote(open('input.svg').read()))"

Python → data URI: 

import base64
def data_uri(path, mime='image/png'):
    with open(path, 'rb') as f:
        b64 = base64.b64encode(f.read()).decode()
    return f'data:{mime};base64,{b64}'

img_uri = data_uri('icon.png')
html = f'<img src="{img_uri}">'

Pillow → data URI without writing a file: 

import io, base64
from PIL import Image
buf = io.BytesIO()
img.save(buf, format='PNG')
uri = f'data:image/png;base64,{base64.b64encode(buf.getvalue()).decode()}'

SVG <image> element with embedded raster: 

<svg viewBox="0 0 400 200">
  <image href="data:image/png;base64,iVBORw0KG..." width="400" height="200"/>
</svg>

Size budget

Base64 inflates by ~33%. Stay under ~200 KB per image inside HTML; above that, prefer external assets unless portability is non-negotiable. SVGs URL-encode smaller than base64 — use urllib.parse.quote instead.

Caveats

  • Data URIs bypass browser cache. Don't use for images that repeat across many pages.
  • Some XML parsers reject very long attribute values.
  • Email clients are inconsistent: Gmail strips data URIs in some contexts, Outlook handles them in HTML body but not signatures.

Anti-patterns

  • Don't exceed ~75 characters per line of body prose. Reading span research backs the 65–75ch cap.
  • Don't animate layout properties (width, margin, top, left). Use transform + opacity only; layout animations trigger reflow every frame.
  • Don't use pure #000 or #fff. Cream + near-black per the OLED palette; pure black/white is harsh on OLED and pure white flares on cream backgrounds.
  • Don't chain more than two CSS custom-property fallbacks. A third fallback means the token system is underspecified — fix the schema, don't paper over it.

Absolute bans

Five escalated anti-patterns — channel-spanning but most visible in web work. Named so the muriel-critique agent can cite them directly and so a human reviewer has a finite list of "stop, don't ship this."

  • gradient-text — background-clip: text with a gradient fill. Unreadable at body sizes (the gradient traverses low-contrast hues by construction, failing 8:1 across its span) and telegraphs designer reached for the default Figma preset. Vary weight/size for hierarchy; keep text one color.

  • side-stripe-borders — card or alert with border-left: 4px solid accent (or any border > 1px on a single side). Identifies the artifact as a Tailwind-CSS-template card at a glance. Use background tint or iconography for the same semantic signal.

  • rounded-rect-with-drop-shadow — default border-radius: 8–12px plus box-shadow: 0 4px 12px rgba(0,0,0,0.1) is the most-recognizable "AI card" aesthetic. If someone can identify the artifact as AI-generated in the first second, it has no distinctiveness. Make a deliberate choice: square corners + no shadow, a pronounced shadow that commits to elevation, or a different grouping signal entirely (proximity, background tint).

  • bounce-easing — cubic-bezier curves with overshoot (elastic, back-out, Material's "playful" namespace). Reads as amateur motion design outside a toy. Use smooth curves — cubic-bezier(0.2, 0, 0.2, 1) or your brand's motion.easing_default token.

  • reflex-fonts — defaulting to the overused web stack without a reason: Inter, DM Sans, Instrument Sans, Geist, Satoshi, or system-ui chosen because it's safe. Fine tools, but everyone reaches for them because Vercel's starters do, which means the artifact looks like every other 2024–2026 landing page. Choose deliberately; if your brand doesn't have a display typeface, consider Helvetica Neue (restrained), Söhne (licensed), Work Sans, IBM Plex, or a serif/slab your brand is willing to commit to. muriel's brand.toml has a [typography] block — use it.

The "absolute bans" escalation and the reflex-fonts concept are borrowed with permission from pbakaus/impeccable (Apache-2.0, Skill 3.1.1 as of May 2026), rephrased in muriel's voice. The underlying insight that a short named-ban list is a stronger instrument than a long positive-rule list is theirs. Impeccable now ships twenty-seven deterministic anti-pattern rules plus twelve LLM-critique rules (a much broader register than what's encoded here) and a standalone npx impeccable detect <path-or-url> CLI that runs regex + headless-Chrome scans with no API key required — a deterministic complement to muriel's vision-model muriel-critique agent. The shortest pipeline for a web artifact: muriel capture → npx impeccable detect → muriel-critique. See HARNESSES.md for cross-harness packaging that mirrors impeccable's eleven-harness matrix.