Typography

Contents

• Fonts that embed (auto-resolve)
• Banned fonts
• Guardrails
• What you do not do without being told
• Finding fonts
• Selection thinking
• Similar-font pairing
• Dark backgrounds
• OpenType features for data

Fonts That Embed (auto-resolve)

These 18 families are the ones the renderer pre-bundles — embedded as local data URIs with no network fetch, so they render offline and deterministically with zero setup, no lint warning, and no fail-closed fetch risk. Write any of them as a font-family and it renders; only the listed weights exist (asking for a weight a family doesn’t ship gives a synthetic/fallback weight, not a real cut). (Any other real Google font still works via the implicit build-time fetch described in the intro — but only these render with none of those caveats.)

⚠ League Gothic and Archivo Black ship weight 400 ONLY — they are already heavy/condensed display faces. Do not request font-weight: 700/900 on them.

Aliases — these common names resolve to an embedded family, so you may safely write them: Helvetica Neue / Helvetica / Arial → Inter · Futura / DIN Alternate / Arial Black → Montserrat · Bebas Neue → League Gothic · Segoe UI → Roboto · Courier New / Courier → JetBrains Mono · Garamond → EB Garamond. (This is why a “safe” Helvetica Neue stack always renders — it maps to embedded Inter.)

Reconciling with the Banned list below: several embedded families (Inter, Roboto, Open Sans, Lato, Nunito, Poppins, Outfit, Playfair Display, EB Garamond) are also on the Banned monoculture list — they render fine but read as generic. The families that are embedded AND not banned — your safe-and-distinctive picks — are: Montserrat, Oswald, League Gothic, Archivo Black, Space Mono, IBM Plex Mono, JetBrains Mono, Source Code Pro, Noto Sans JP. Reach for these (or a non-bundled font you’ve confirmed via the Finding-Fonts step). A non-bundled name isn’t guaranteed-broken — a real Google font is auto-fetched and embedded — but it carries a lint warning and a fail-closed fetch in cloud renders, so for anything that must render predictably, embed it yourself via @font-face (see “Finding Fonts”) rather than relying on the implicit fetch.

Banned

Training-data defaults that every LLM reaches for. These produce monoculture across compositions.

Inter, Roboto, Open Sans, Noto Sans, Arimo, Lato, Source Sans, PT Sans, Nunito, Poppins, Outfit, Sora, Playfair Display, Cormorant Garamond, Bodoni Moda, EB Garamond, Cinzel, Prata, Syne

Syne in particular is the most overused “distinctive” display font. It is an instant AI design tell.

Guardrails

You know these rules but you violate them. Stop.

• Don’t pair two sans-serifs. You do this constantly — one for headlines, one for body. Cross the boundary: serif + sans, or sans + mono.
• One expressive font per scene. You pick two interesting fonts trying to make it “better.” One performs, one recedes.
• Weight contrast must be extreme. You default to 400 vs 700. Video needs 300 vs 900. The difference must be visible in motion at a glance.
• Video sizes, not web sizes. Full-screen viewing (YouTube / website embed): body 20px minimum, headlines 60px+, data labels 16px. In-feed viewing (destination = X / LinkedIn / Instagram feed — hyperframes/references/brief-contract.md § 2): the video plays small inside a scrolling feed, so scale up — body ≥32px, headlines ≥90px, data labels ≥24px (first-pass values; calibrate against real renders). You will try to use 14px. Don’t.

What You Don’t Do Without Being Told

• Tension should mean something. Don’t pattern-match pairings. Ask WHY these two fonts disagree. The pairing should embody the content’s contradiction — mechanical vs human, public vs private, institutional vs personal. If you can’t articulate the tension, it’s arbitrary.
• Register switching. Assign different fonts to different communicative modes — one voice for statements, another for data, another for attribution. Not hierarchy on a page. Voices in a conversation.
• Tension can live inside a single font. A font that looks familiar but is secretly strange creates tension with the viewer’s expectations, not with another font.
• One variable changed = dramatic contrast. Same letterforms, monospaced vs proportional. Same family at different optical sizes. Changing only rhythm while everything else stays constant.
• Double personality works. Two expressive fonts can coexist if they share an attitude (both irreverent, both precise) even when their forms are completely different.
• Time is hierarchy. The first element to appear is the most important. In video, sequence replaces position.
• Motion is typography. How a word enters carries as much meaning as the font. A 0.1s slam vs a 2s fade — same font, completely different message.
• Fixed reading time. 3 seconds on screen = must be readable in 2. Fewer words, larger type.
• Tracking tighter than web. -0.03em to -0.05em on display sizes. Video encoding compresses letter detail.

Finding Fonts

Don’t default to what you know. If the content is luxury, a grotesque sans might create more tension than the expected Didone serif. Decide the register first, then search.

Save this script to /tmp/fontquery.py and run with curl -s 'https://fonts.google.com/metadata/fonts' > /tmp/gfonts.json && python3 /tmp/fontquery.py /tmp/gfonts.json:

Five categories: trending sans, trending serif, monospace, impact/condensed, script/handwriting. All dynamically filtered from Google Fonts metadata — no hardcoded font names. Cross classification boundaries when pairing.

Selection Thinking

Don’t pick fonts by category reflex (editorial → serif, tech → mono, modern → geometric sans). That’s pattern matching, not design.

1. Name the register. What voice is the content speaking in? Institutional authority? Personal confession? Technical precision? Casual irreverence? The register narrows the field more than the category.
2. Think physically. Imagine the font as a physical object the brand could ship — a museum exhibit caption, a hand-painted shop sign, a 1970s mainframe terminal manual, a fabric label inside a coat, a children’s book printed on cheap newsprint, a tax form. Whichever physical object fits the register is pointing at the right kind of typeface.
3. Reject your first instinct. The first font that feels right is usually your training-data default for that register. If you picked it last time too, find something else.
4. Cross-check the assumption. An editorial brief does NOT need a serif. A technical brief does NOT need a sans. A children’s product does NOT need a rounded display font. The most distinctive choice often contradicts the category expectation.

Similar-Font Pairing

Never pair two fonts that are similar but not identical — two geometric sans-serifs, two transitional serifs, two humanist sans. They create visual friction without clear hierarchy. The viewer senses something is “off” but can’t articulate it. Either use one font at two weights, or pair fonts that contrast on multiple axes: serif + sans, condensed + wide, geometric + humanist.

Dark Backgrounds

Light text on dark backgrounds creates two optical illusions you need to compensate for:

• Increased apparent weight. Light-on-dark reads heavier than dark-on-light at the same font-weight. Use 350 instead of 400 for body text. Headlines are less affected because size compensates.
• Tighter apparent spacing. Light halos around letterforms reduce perceived gaps. Increase line-height by 0.05-0.1 beyond your light-background value. For display sizes, add 0.01em letter-spacing to counteract.

OpenType Features for Data

Most fonts ship with OpenType features that are off by default. Turn them on for data compositions:

tabular-nums is essential any time numbers are stacked vertically — stat callouts, timers, scoreboards, data tables. Without it, digits have proportional widths and columns don’t align.