Ratio-based type scales with CSS and Tailwind tokens, minimum sizes, letter-spacing, caps labels and weight on dark backgrounds.
$ npx skills add https://craftwork.design/skills/modular-scale-typography -a claude-code -y
Installs into .claude/skills/modular-scale-typography. Add -g for every project.
SKILL.md
Typography feels cohesive when all font sizes are related to each other through a single mathematical ratio. Without a scale, sizes get picked arbitrarily and the result feels visually noisy — headings that don't contrast enough, body text too close in size to captions, labels that blend into content.
What Is a Modular Scale
A modular scale starts from a base size and multiplies or divides by a ratio to generate every size in the system.
size(n) = base × ratio^n
Every size is thus a deliberate step away from the base — not a guess.
Default recommendation:1.25 (Major Third) — enough contrast between steps to feel intentional without being theatrical.
Generating a Scale
Recover an existing scale, don't reverse-engineer it by hand (dembrandt engine, optional). If a brand already has type on the web, get_typography returns the real font sizes, weights, and line-heights computed off the live DOM — infer the underlying ratio from those, then regularise it with the method below, instead of guessing which sizes were intended. See extract-design.
Starting from base = 16px, ratio 1.25:
Step
Formula
Value
Rounded
Role
-2
16 ÷ 1.25²
10.24px
10px
Caption, label-xs
-1
16 ÷ 1.25
12.80px
13px
Label, small
0
16
16px
16px
Body (base)
+1
16 × 1.25
20px
20px
Body-lg, lead
+2
16 × 1.25²
25px
25px
H4
+3
16 × 1.25³
31.25px
31px
H3
+4
16 × 1.25⁴
39.06px
39px
H2
+5
16 × 1.25⁵
48.83px
49px
H1
+6
16 × 1.25⁶
61.04px
61px
Display
Round to whole pixels or rem — the ratio provides the intent, exact pixel rounding is fine.
Without a scale, designers and developers pick sizes by eye or habit (14px, 16px, 18px, 24px, 32px, 48px). These feel subtly wrong because the intervals are uneven — the jump from 14→16 is small, 32→48 is large, and there is no underlying logic tying them together. The eye senses the inconsistency even when the viewer cannot name it.
With a modular scale, every size step carries the same visual weight of change. Hierarchy reads clearly because each level is a proportional step away from the next, not an arbitrary gap.
Minimum Font Size
Body text base: 16px minimum. This is the browser default for good reason — it is the threshold below which reading comfort drops significantly, especially on screens.
16px — standard body text, the default base
14px — acceptable for secondary UI text (labels, captions, metadata) used sparingly
Below 14px — do not use. Even at high DPI, sub-14px text fails WCAG contrast requirements for normal text and creates accessibility issues.
In the modular scale, this means the base (step 0) should be 16px, and negative steps (step -1, step -2) should be used only for genuinely secondary content — never for body copy or primary labels.
The 1% heuristic. Sub-16px text is the rare exception (target: under ~1% of a page, never below 14px). Nobody counts characters, so apply it as a role test:
May go below 16px — a fixed whitelist of glanced-at roles: timestamps, captions, table metadata, helper text, fine print, badge labels.
Never — body copy, primary labels, list item titles, anything actually read to do the task.
Check: scan one screen, count distinct sub-16px roles. Two or three from the whitelist is healthy; five or more — or any reading content — means you've over-shrunk. That's layout density, not a type problem: cut what's shown rather than shrink the type to fit.
Type Rendering Details
Size and ratio set the structure; these details determine whether the type actually reads well on screen.
Letter Spacing
Body and default text: keep tracking at 0. Adding it to running text slows reading and makes the type feel loose.
Uppercase and small labels: the one place tracking helps, since capitals are visually tight. Cap it at 0.04em, and reach for that only when the label genuinely needs air.
Pattern: zero on lowercase body, at most a hair (≤ 0.04em) on uppercase labels — never a blanket value. Over-tracking reads as dated, not premium.
Optical Gaps Around Caps Labels
An all-caps or descender-free label leaves the bottom of its line box empty. Equal margins above and below it measure equal in the inspector and read larger below. Measure gaps from the visible glyphs, not the box, and shave a few pixels from the lower gap until the two look equal. This is what people mean when spacing is "off" but they cannot say why.
Weight on Dark Backgrounds
Light text on dark appears optically thinner — a halation effect where bright type bleeds into the dark field. Step up one weight to compensate: where regular (400) works on light, use medium or semibold for the equivalent text on dark. This keeps perceived weight consistent across modes instead of dark-mode text looking frail.
Monospace
Monospace is for technical content where character alignment matters — code, IDs, numeric tables, diffs. Don't use it as a default UI typeface, and avoid monospace + uppercase (even widths plus tall capitals are hard to scan).
Numbers That Change in Place
A counter, a countdown, a price that updates, a column of figures — proportional digits give each number a different width, so the text jitters sideways on every tick and columns fail to line up. Switch to tabular figures anywhere digits update or stack; leave body text proportional, where tabular digits look gappy.
Headings are the place to control the break. text-wrap: balance evens the lines of a short heading and stops the one-word last line; text-wrap: pretty only prevents the orphan and is the cheaper choice for paragraphs. Balance is capped at a handful of lines by the browser, so it is a heading tool, not a body tool.
Set it once, in the stylesheet, for every element of that kind. A
text-balance utility sprinkled on one component holds for that component and
silently stops holding on the next page someone builds.
All three browsers balance; only Chrome and Safari do pretty. Neither is
load-bearing. Balance stops at six lines (Chrome) or ten (Firefox), so never on
body copy. Titles, toasts, tooltips, captions and wrapping buttons strand words
too: add them by class to the same rule. Balance moves breaks, not the box, so
cap a left-aligned heading with max-width in ch first.
Never break a line by hand with <br> to fix a widow. It is correct at exactly
one viewport width and wrong at every other. Per-component wrapping is a
deliberate exception a designer asks for, not a default you reach for.
Bind a word to what follows it with a non-breaking space — 10 kg, Figure 3, a name and its title. A unit stranded alone on the next line reads as a typo.
Text That Does Not Fit Its Slot
text-wrap: balance fixes a heading that breaks awkwardly. It does not fix text
that is simply too long for the space it was put in, and reaching for it there
hides the real problem for one viewport width and returns at the next.
When a line orphans a word, or a caption wraps to three lines beside a title,
there are exactly three honest fixes:
Shorten the text. Usually the right one. A slot sized for a label was not
asking for a sentence.
Drop it a step on the scale. Supporting text beside a title belongs a step
or two below it, not at the same size.
Change the parent layout. When the text is already as short as it can say
what it means, the container is the problem. Fewer grid columns at that
breakpoint, a wider column, its own line, a different position. Six tiles
across a 1280px viewport leaves each about 190px, and no rewording makes a
number, a label and a caption fit that; three across at the same width does.
The order matters. Reach for the layout when the first two have run out, not
before: widening a container to rescue a sentence that should have been three
words moves the same problem to the next breakpoint.
What is not a fix: leaving it, or nudging the container until it happens to fit
the width you are looking at.
A legend is not a sentence. "Bar height is the total, the filled part is how
many completed" is thirteen words in a header sized for three. It wraps, and it
strands a word on the last line. Two swatches and one word each carry the same
meaning and cannot wrap badly:
■ completed □ total
The same applies to axis labels, table headers, chip text, and any caption
sitting next to something larger. If the explanation genuinely needs a sentence,
it belongs under the element rather than beside it.
Two Lines in One Slot
A label with an explanation under it is two jobs, and they have to look like two
jobs. Given the same size, weight and colour they merge into one block of grey
that reads as neither:
<!-- wrong: identical styling, both clipped -->
<div class="w-36">
<div class="text-sm font-medium text-gray-500 truncate">Imported</div>
<div class="text-sm font-medium text-gray-500 truncate">rows without a matching account are skipped</div>
</div>
Separate them by weight and ink, not by size — this is the exception to
"different jobs step apart" above: both lines already sit near the 14px floor,
so dropping the second a step pushes it under the minimum, and two steps apart
is barely visible at these sizes anyway. Semibold primary over medium muted is
enough, and the muted colour needs a floor of its own: pair the palette's
faintest legible colour with the lightest weight (400) and it reads as
illegible even where it passes WCAG contrast on colour alone, because contrast
math is colour-only and ignores stroke weight. Give any muted/secondary text
token a minimum weight of 500 wherever it is used, not just here.
Do not truncate an explanation. A name survives clipping because a reader
recognises it from its first characters. A sentence does not: what is cut is the
part that carried the meaning, and moving it into a title attribute hides it
from touch, from keyboards, and from anyone who does not think to hover — this
is why the title-attribute escape hatch under Line Clamping below is scoped to
supporting descriptions the user does not need to act on, never to an
explanation that carries the meaning of what it is attached to. Truncate
identifiers. Let explanations wrap, shorten them, or give the column more room.
Type Scale by Page Context
Landing pages and marketing surfaces benefit from large, expressive type — steps +4 to +6 for headlines create drama and brand presence.
Feature pages and application UI should use a more controlled range — steps +2 to +3 for headings, with body text at step 0. Oversized headings inside a functional UI distract from the content and make the layout feel unbalanced.
Match the ratio and scale usage to the purpose of the surface, not just the brand.
Four steps per view. Within one view and one context, a panel, an overlay, a page region, use no more than four steps of the scale. Four is not a taste threshold, it is the number of type roles a single view has: title, body, label, metadata. The cap follows from the roles, so a view that seems to need a fifth step is really carrying a fifth role nobody has named, and naming it usually reveals it belongs to a different view or to a group the layout has not drawn yet. Every additional size is also a claim about hierarchy, and past four the eye stops ranking them, so adjacent steps read as mistakes rather than levels. This is the same argument the Rule of Three below makes about heading levels; this one is about type sizes on screen, and a view can hit its four-size limit with a single heading level in it.
Heading Hierarchy and Page Complexity
A successful heading scale uses more than just font size to distinguish levels. It also respects the cognitive limits of the page.
Tools for Differentiation
If headings only differ by small increments of size, they become hard to distinguish at a glance. Use these tools to create a more meaningful scale:
Capitalization: Use uppercase (text-transform: uppercase) for small, lower-level headings (H4–H5) to give them visual weight without needing large sizes.
Letter Spacing: When using uppercase or bold headings, add at most a hair of letter-spacing (≤ 0.04em) — and only when the heading genuinely needs the air. Keep it subtle; over-tracking reads as dated, not premium.
Color: Use your brand primary colour or a slightly muted grey for secondary headings to differentiate them from the main black/dark-grey text.
Style: Use italics or subtle underlines for supplementary or metadata-style headings.
The Rule of Three (H1–H3)
Most well-designed pages require only three levels of heading hierarchy (H1, H2, H3).
Simplicity: H1–H3 is enough to cover the page title, section titles, and sub-sections.
H4–H6 is a complexity smell, not a typography problem. If you find yourself reaching for H4, H5, or H6, the page is trying to do too much. Adding smaller heading levels only hides the symptom. Read it as a signal and make a structural decision instead:
Split — the page is really two features, or an in-page sub-page. Break it across pages or a modular structure (sidebar, tabs, master-detail).
Remove — the deep section may not earn its place at all.
Simplify — flatten the sub-hierarchy so it fits within H1–H3.
Hide behind an opening/closing element — push secondary content into a modal, accordion, or panel (see information-architecture).
A page that needs 6 levels of headings is a page that most users will stop reading.
Reading Comfort and Editorial Patterns
Typography is not just about size — it is about the rhythm and structure of the content.
Line Length (Measure)
For optimal reading comfort, keep body text between 45–75 characters per line (approx. 500–700px).
Lines that are too long make it hard for the eye to find the start of the next line.
Lines that are too short break the reading rhythm and create distracting "rags."
Measure, Leading and Size Move Together
Line length, line-height (leading) and font size are not three independent knobs — they are one system. This is a foundational typographic principle (Bringhurst's The Elements of Typographic Style) repeatedly validated on digital platforms, and it underpins the line-height ramps in the Material and Apple HIG type scales.
The longer the measure, the more leading the eye needs to track from the end of one line back to the start of the next. A tight line-height that reads fine on a narrow column becomes tiring on a wide one.
Measure
Recommended body line-height
Narrow (~45ch)
1.4
Comfortable (55–66ch)
1.5
Wide (~75ch)
1.6–1.7
Practical rules:
Body text: line-height 1.4–1.7, never below 1.4 for multi-line copy. WCAG SC 1.4.12 also requires text to stay readable when users override line-height to at least 1.5, so design with that headroom.
Headings: large type needs less leading — tighten to 1.1–1.25 as size grows, or the lines drift apart and stop reading as one unit. Leading and size are inversely related.
Fix the cause, not the symptom: if a line of body text feels hard to read, widen the leading or narrow the measure — don't just shrink the font. The three move together.
Line Clamping
In grids, cards, or lists with unpredictable content lengths, clamp text to keep a consistent visual rhythm and equal-height cards. Limit descriptions to 2–3 lines so all cards in a row stay the same height.
Multi-line clamp. Pair the standard line-clamp with the -webkit- fallback — the legacy -webkit-box form is still required for full browser support, so keep both.
Inside flex or grid. Truncation silently fails when the item can't shrink below its content width. Add min-width: 0 (or min-inline-size: 0) to the clamped child so it is allowed to narrow.
Accessibility — never lose the content. Clamping hides text visually only; the full string stays in the DOM and is read in full by screen readers, but a sighted user can no longer see it. Always give them a path to the rest:
Provide a title attribute (or accessible tooltip) carrying the full text, or
Link to a detail view / "Read more" where the complete content lives.
This escape hatch is for supporting descriptions the user does not need to act
on. It does not extend to an explanation whose meaning is in the clipped part —
see "Do not truncate an explanation" above.
Never clamp text that the user must read to act (prices, errors, legal copy, primary instructions) — clamp only supporting descriptions where truncation is safe.
Fixed-size pages. In a document built from fixed pages with overflow: hidden (print, PDF, A4 preview), text appended to a page the renderer already filled from measured content lands below the fold and prints on no page at all, with no error. Generated or authored copy gets a page of its own with a box bounded above the footer, and the clamp is applied in the browser where the wrapping happens. A character budget cannot predict where a page ends; only the layout engine knows.
Editorial Hierarchy
Use specific typographic roles to provide context and guide the user through the story:
Role
Visual Treatment
Purpose
Pre-title (Eyebrow)
Small (12–13px), often all-caps, subtle letter-spacing (≤ 0.04em), muted colour
Provides context or category without distracting from the main heading
Heading
Large, bold, modular scale step +3 to +5
The primary hook or subject
Ingress (Lead text)
Larger than body (step +1), slightly bolder or higher line-height
Summarises the core value; bridging the heading and the body copy
Body
Base size (16px), regular weight, comfortable line-height (1.5)
The primary reading experience
Wording and Voice
Use active voice. "Get started" instead of "Getting started is easy."
Be punchy. Use clear, descriptive labels that promise a result.
Consistency. Use the same terms for the same actions throughout the product.
Responsive Type Scale
The scale compresses on smaller viewports by tightening the ratio, not by manually overriding individual sizes. The floor (body, labels) stays stable — readability has a hard minimum. The ceiling (H1, display) shrinks significantly. Every heading level scales proportionally because the ratio changes; hierarchy stays internally coherent.
Rule: use a wider ratio on desktop (more drama), a tighter ratio on mobile (more restraint). The base stays the same. The top end gives way.
Desktop → wider ratio (e.g. 1.25–1.333): large contrast between H1 and body
Tablet → moderate ratio (e.g. 1.200): scale pulls in
Mobile → tight ratio (e.g. 1.125): H1 is notably smaller, body is unchanged
Never shrink the scale from the bottom. Body text at 16px is already a floor — compression always comes from the top.
Fluid Type with clamp()
Instead of stepping sizes at fixed breakpoints, let the top end scale smoothly between a floor and a ceiling using CSS clamp(MIN, PREFERRED, MAX). The viewport-relative middle term does the scaling; the floor and ceiling stop it from ever getting too small or too large.
The MIN is the mobile size, the MAX is the desktop size — the same floor/ceiling thinking as the stepped scale, just interpolated.
Apply clamp() to the top of the scale (headings, display). Keep body and labels at a fixed size — readability has a hard minimum, so the floor must not move.
Include a rem term in the preferred value (e.g. 1.5rem + 3vw, not 3vw alone) so the text still responds to user zoom and root font-size — a pure vw value breaks zoom accessibility.
Adjacent Fluid Steps Can Converge
If a design system fluidizes several small steps anyway (text-fluid-xs, text-fluid-sm, and so on) rather than keeping them fixed, two adjacent clamped sizes are not guaranteed to stay apart across the viewport range. Each clamp() interpolates on its own slope between its own floor and ceiling, so two steps with different vw coefficients can land close together — or land on the exact same computed value — at some width in between, even though they look correctly spaced at the narrowest and widest extremes. Checking only the two endpoints hides this: the gap that matters is the minimum gap across the whole range, not the gap at either end.
/* Looks fine at 375px and 1440px. At ~860px both resolve to ~14.9px. */
--text-fluid-xs: clamp(0.75rem, 0.65rem + 0.4vw, 0.875rem);
--text-fluid-sm: clamp(0.8125rem, 0.7rem + 0.45vw, 0.9375rem);
Don't fluidize adjacent small steps at all. This is the same "keep body and labels fixed" rule above, stated for the case it actually gets violated: a caption sitting directly beside body text, or two label-scale steps next to each other, should use fixed sizes precisely so their difference can't collapse. Reserve clamp() for the top of the scale, where one heading rarely sits pixel-adjacent to another size on the same line.
If a small step must stay fluid, verify the gap across the range, not just at MIN and MAX: sample the pair's computed values at a handful of intermediate widths (or plot both clamp() expressions) and confirm neither crosses nor closes to within a couple of px anywhere in between, not only at the breakpoints you happened to check.
Prefer a shared coefficient. Giving adjacent fluid steps proportionally related vw coefficients (e.g. derived from the same modular ratio) keeps their curves parallel instead of letting independently-chosen slopes cross.
Prefer 16px for All Reading Text
Hints, metadata, captions and HUD labels are read too, so they default to the same 16px as body copy. Build hierarchy with weight, colour and spacing first and go down the scale only when those are spent; secondary text never drops below 14px. Each step below base is a deliberate exception, not the default for anything called secondary.
Review Checklist
Are all font sizes derived from a single base + ratio?
Is the base size 16px or larger?
Is 14px used only for secondary/metadata text, never for body copy?
Is nothing below 14px used anywhere in the UI?
Counting distinct sub-16px text roles on a screen, are there only a few (≤3) and all from the secondary whitelist — never reading content?
Is body letter-spacing 0, with uppercase-label tracking capped at 0.04em and used only when air is needed?
Does light-on-dark text use a slightly heavier cut to compensate for halation?
Are gaps around all-caps labels balanced by the visible glyphs, not by the line box?
Is monospace scoped to technical content, never used as a default or with uppercase?
Is there at least 3–4 distinct steps between body text and the largest heading?
Are adjacent steps (e.g. body vs. label) different enough to be distinguishable at a glance?
Are font size tokens named by role (--text-body, --text-h1) or step (--text-base, --text-2xl), not by raw pixel value?
Does the chosen ratio suit the UI density? (tight ratio for data-heavy UIs, wider ratio for marketing)
Is body text line length between 45–75 characters?
Does any supporting text beside a title wrap, or strand a word on its own line? Shorten it, step it down the scale, or give it room; never leave it.
Is text-wrap: balance set once, for headings and short-copy classes, and off body copy?
Are legends, axis labels and chip text written as labels rather than sentences?
Where text still does not fit after shortening, has the column count at that breakpoint been reconsidered, rather than the text squeezed further?
Where a label and its explanation stack in one slot, do they differ in weight and ink rather than being two identical grey lines?
Is truncation used only on identifiers, never on an explanation whose meaning is in the part being cut?
Does body line-height scale with the measure (≈1.4 narrow → ≈1.6+ wide), staying ≥1.4 and leaving 1.5 override headroom for WCAG 1.4.12?
Is heading leading tightened (≈1.1–1.25) as size grows, so large type still reads as one unit?
Is line-clamping used to keep grid/card layouts consistent?
Does clamped text keep a path to the full content (title/tooltip or detail view), and is must-read content never clamped?
If fluid type (clamp()) is used, is it applied only to the top of the scale, with a rem-based preferred term so zoom still works?
If two adjacent steps are both fluid (e.g. xs/sm), do they stay visibly distinct across the whole viewport range, not just at the MIN/MAX extremes?
Are editorial roles like pre-titles and lead text used to improve scannability?
Are headings differentiated by more than just size (e.g., color, case, spacing)?
Is the heading hierarchy limited to H1–H3 per view where possible?
Is reading text, including hints and metadata, 16px by default, with hierarchy carried by weight and colour before size, and nothing below 14px?
Common Anti-Patterns
Anti-pattern
Problem
Fix
Sizes like 14, 15, 16, 17px used side by side
Steps too small to read as distinct levels
Use at minimum a 1.125 ratio so each step is perceptible
Arbitrary sizes with no relationship (13, 18, 27, 36px)
No underlying logic — hierarchy feels accidental
Regenerate from a single base and ratio
Pixel values hard-coded in components instead of tokens
Scale changes require hunting through every file
Define once as CSS custom properties or design tokens
Same scale used for display headings and dense data tables
One ratio rarely serves both extremes well
Use a tighter ratio (1.125) for data, wider (1.25–1.333) for marketing contexts
Letter-spacing added to running body text
Loosens the type and slows reading
Keep body tracking at 0; add at most 0.04em to uppercase labels when air is needed
Regular-weight white text on a dark background
Halation makes it look thin and frail
Step up one weight (medium/semibold) for light-on-dark text
Monospace as a default UI font, or monospace + uppercase
Hard to scan, reads as "unstyled"
Scope monospace to code/IDs/numeric data only
Adjacent small steps both set with independent clamp() values
Their curves can converge or cross mid-viewport, even when the MIN/MAX extremes look fine
Keep adjacent small/label steps at a fixed size; reserve clamp() for the top of the scale
