TL;DR
We replaced Seed's imperative, per-block-type spacing rules with a single .hm-prose class that scales typography the way the browser was designed to — relative units on real semantic structure instead of hard-coded pixel values scattered across wrappers. Inline styling rides the semantic tag (<strong>, <em>, <code>, <a>); section rhythm lives on the block wrapper and is keyed off stable data-node-type / data-content-type attributes. The result: documents that breathe naturally, fewer stylesheets to keep in sync, and the same rhythm in the desktop app, on the web (including SSR), inside embeds, and in comments.
The problem
Until now, Seed's block spacing was hand-coded: every heading level, every media type, and every "what comes before what" combination had its own CSS rule. These rules were written to a pixel grid and spread across three stylesheets (editor.css, Block.module.css, blocks-content.css) plus a manual mirror for the web SSR view in apps/web/app/styles.css.
It worked, but it felt forced:
An H1 always added 48px above itself, regardless of the size of the text around it.
Headings used one rule, images used another, code blocks a third, and the rules fought each other via sibling combinator specificity (:has(> [data-content-type='heading']) beats the default, which beats the first-child reset…).
Changing one spacing decision meant finding and updating it in four places.
The SSR stylesheet was a manual reimplementation of the editor stylesheet — every refactor risked drift.
Before: section markers, media, and code each declared their own wrapper-level margin. Nothing scaled with text.
The principle
Great typography on the web has a settled answer for this, popularized by Tailwind Typography and long used by publications like Medium, Substack, and the NYT:
Semantic HTML is the source of truth. <strong>, <em>, <h1>, <blockquote>, <pre> — not <span class="font-bold"> or <div class="mt-12">. Inline marks, colors, fonts, and heading scale all live on the semantic tag.
Section rhythm lives on the block wrapper, keyed by a stable data attribute. data-node-type="blockNode" and data-content-type="heading" survive CSS-modules hashing and edit-vs-SSR DOM differences — they're the one safe hook we have for saying "give an H1 a section-marker gap above it."
Units are relative. em on tag-level rules (heading scale, inline code padding), rem on wrapper-level section gaps (so the gap above an H1 doesn't depend on the H1's own font-size). No manual mobile override needed.
Scope the typography to a single class on the container. Layer your overrides descendant-style so they pass through any number of structural wrappers untouched. .hm-prose for documents, .hm-prose.is-comment for comments — one class flip swaps the whole voice.
After: one .hm-prose class. Headings, paragraphs, lists, code, quotes — each owns its own rhythm in em units.
What we changed
One class, four renderers. .hm-prose (defined in packages/ui/src/hm-prose.css) is applied to the draft editor, the read-only viewer, the embed wrapper, the comment body, and the web SSR placeholder. Same rhythm everywhere.
@tailwindcss/typography as the foundation. The plugin gives us tested, em-based defaults wrapped in :where(…) selectors at zero specificity. .hm-prose layers our font stack, link treatment, code pill, blockquote border, heading scale, and comment variant on top — whenever we diverge, we win cleanly.
Selectors target data-node-type / data-content-type attributes, not CSS-module class names. The editor package's Block.module.css hashes classes like .blockNode to _blockNode_abc123 at build time, so any global rule that tries to target .blockNode silently misses. The data attributes — emitted verbatim by BlockNode.renderHTML and mirrored by the SSR renderer — are the one safe hook.
Deleted the imperative :has() + --space-* spacing rules. The block of rules in editor.css / Block.module.css / blocks-content.css that hard-coded 48px above every H1, 32px above every media block, and so on, is gone. The 8px-grid --space-xs..--space-xl tokens are removed across all three stylesheets. Prose em/rem defaults replaced every use.
Moved section-marker rhythm to the wrapper. Heading top-margins live on the .blockNode wrapper via :has(> [data-content-type='heading'][data-level='1']), not on the <h1> element. This makes the first-child reset (:first-child { margin-top: 0 }) a single-selector win regardless of whether the heading is wrapped in editor DOM or emitted directly in SSR. Nested first-children get a small 0.25em separator so they're clearly distinct from their parent's content.
Inline code is a real pill. .hm-prose code now has a visible tinted background, monospace stack, 0.875em size, and an accent color — slate-red in light mode, soft pink in dark mode. Both meet WCAG AA on the corresponding muted background.
Blockquotes sit flush with the parent paragraph column. The left border no longer indents to 1.5em — only the 15px padding between the border and the quoted text remains. Reads more like a side-annotation than a nested card.
Removed every font-family: … !important override in the editor. editor.module.css had .defaultStyles h1..h6 forcing Inter !important; font-weight: 400; font-size: 18px (that's why your H1s looked like body text) and .draft-editor .defaultStyles forcing Georgia serif — plus a leftover background-color: red debug. All gone. Font-family and heading scale are now owned by .hm-prose; the module only holds structural resets.
Purged the Mantine font-family leak. BlockNote's Mantine theme was emitting .mantine-xxxx .ProseMirror { font-family: Inter, … } via Emotion's CSS-in-JS. Vite HMR caches Emotion stylesheets across editor remounts, so right after a markdown import the draft editor rendered in sans-serif until a full reload rebuilt the cache. Removing the fontFamily declaration from the Mantine Editor.root style means Emotion emits no font-family rule at all, and .hm-prose's serif stack always wins.
One baseline, four size modifiers. .hm-prose sets 1.125rem (18px) as the document body size; .is-comment drops to 1rem (16px). For surface-specific overrides, apply .hm-prose-sm (0.875rem), .hm-prose-base (1rem), .hm-prose-lg (1.125rem), or .hm-prose-xl (1.25rem) — em-based rhythm scales automatically.
Made the editor wrappers structurally transparent. .blockNode, .blockContent, and .blockChildren still exist (ProseMirror decorations need them), but they no longer carry margins of their own. Nothing between the container and the semantic tag fights the prose rhythm.
Purged the vestigial renderer. The .blocknode-content / [data-block-type=...] selectors in blocks-content.css targeted a renderer that no live component emits anymore. Deleted. The file shrank from ~400 lines of content-styling to ~30 lines of embed chrome.
Shrunk the web SSR mirror. The manual reimplementation of the editor's Block CSS in apps/web/app/styles.css went from ~470 lines to ~150 lines. Prose owns body typography; the SSR CSS owns only wrapper structure, list-item display, code-block chrome, colors, alignment, and SSR-only card fallbacks.
Same document, same blocks. Left: before. Right: after.
Font stack
We also swapped in a system-font stack:
Document body: ui-serif, Georgia, Cambria, "Times New Roman", Times, serif. On macOS/iOS this picks up New York; on Windows it picks up Cambria; elsewhere it falls back to Georgia. Zero network cost, no flash of unstyled text.
UI and comments: ui-sans-serif, system-ui, -apple-system, "Segoe UI", … — the browser's native sans.
Code: unchanged — ui-monospace, SFMono-Regular, "SF Mono", Menlo, ….
This removes every font-family: … !important rule from the content stylesheets. The UI picks up the native look on each OS, and comments stay visually distinct from docs because .hm-prose.is-comment swaps the font stack.
Same page on three operating systems. Native on each, no webfont loader required.
What did not change
Block schema and document format. Nothing in the hypermedia data model changed. Existing documents render identically in shape — only their typographic rhythm improves.
Editing behavior. Keyboard shortcuts, block commands, drag handles, hover actions, and the slash menu are untouched.
Syntax highlighting for code blocks. Still lowlight + highlight.js with the same token theme. Shiki migration is a separate follow-up.
What's next
Shiki-based code blocks. Move to VS-Code-grade tokenization with JSON themes and native SSR support. (Own PR.)
TipTap-free read-only renderer for the web. Today the web app dynamically loads the full editor even to display a document. A dedicated, semantic-HTML-only renderer would ship less JS, render faster, and let us delete the SSR mirror entirely. (Own PR.)
Self-hosted editorial fonts (Source Serif 4 / Newsreader pair) if we decide Seed wants a distinctive brand voice in the body text.
How to try it
git checkout simplify-editor-dom
pnpm install (pulls in @tailwindcss/typography)
pnpm --filter desktop dev — open a rich doc with headings, lists, code, media, and a quote.
pnpm --filter web dev — visit the same doc on the web to confirm SSR and hydrated output match.
A companion stress-test document at packages/editor/docs/stress-test.md exercises every block type — headings, inline marks, lists (with nesting), blockquotes, three code blocks, horizontal rule, and placeholder sections for each media and embed type. Import it via seed-cli draft create -f …/stress-test.md or paste into a new draft to review the rhythm in one scroll.
Toggle Cmd/Ctrl+Shift+D in the editor to show the block-boundary debug overlay — useful for seeing that wrappers now carry no margin of their own.
Do you like what you are reading? Subscribe to receive updates.
Unsubscribe anytime