Keeps buttons, inputs, badges and pills consistent in radius, border, height, shadow and color, so they speak one interaction language.
$ npx skills add https://craftwork.design/skills/component-family-consistency -a claude-code -y
Installs into .claude/skills/component-family-consistency. Add -g for every project.
SKILL.md
Every interactive component in a product — buttons, inputs, selects, checkboxes, radio buttons, pills, badges, tags, calendars, date pickers, sliders, toggles — belongs to the same visual family. They share a common design DNA. A user should be able to look at any component and feel that it belongs to the same product as every other component.
When components are designed in isolation without shared tokens, the product feels assembled from parts rather than built as a whole.
Reuse Before You Build a New Component
Before creating any component, audit what already exists — a new-from-scratch component is another mouth to feed: another entry in the family that must stay consistent (radius, height, states, motion) and another thing to maintain. Building fresh should be the last resort, not the first move. Work down this order:
Is there already a component that does this? Use it as-is. If it almost fits, extend it with a prop or variant rather than cloning it — one flexible Button beats PrimaryButton, BigButton, and CtaButton living in parallel.
Is there something close in the codebase you can generalise? Often a one-off was built inline for a single screen. If a small change would make it generic — lift it into the shared library, parameterise the hard-coded bits (label, colour, size via props/tokens), drop the screen-specific assumptions — do that instead of writing a second near-identical thing.
Only build new when nothing existing fits and nothing can be reasonably generalised — and when you do, build it from the shared DNA below so it joins the family cleanly.
Parallel one-offs — three near-identical buttons, two cards with different radius — are how a design system drifts. Before adding a component, ask: does this exist, or is it one refactor from existing?
The inverse also holds: a visual treatment that appears independently in 2–3 places has earned promotion — name it and make it a shared component or token before a fourth copy appears. And when pages from different design eras disagree, migrate old toward new: the newest components are the best evidence of current intent, but confirm before deprecating a style — see generate-ui-from-brand for the consolidation pass.
Find the inconsistency automatically (dembrandt engine, optional). Spotting where a live product has already diverged — five near-identical button radii, three greys that should be one — is tedious by eye. get_findings runs a design-system lint over a real extraction and reports consistency and duplication issues to consolidate. (And compute_drift scores how far two extractions have drifted apart — e.g. this product vs. its reference, or before vs. after a cleanup.) Use them to audit an existing product before deciding what to reuse. See extract-design.
The Shared DNA
Define these tokens once. Every component inherits from them.
Border-Radius
All interactive components use the same base radius token. Variations are derived, not invented.
A button and an input on the same form must have the same radius. A pill is always --radius-full. A badge is --radius-sm or --radius-full depending on brand tone — but consistent across all badges.
Nested corners are concentric. When one rounded box sits inside another, the outer radius equals the inner radius plus the gap between them. A card with 12px padding around an 8px button needs 20px, not 12px. Get this wrong and the corners run at different curvatures a few pixels apart — nobody names it, everybody sees it.
Derive it with calc() rather than hardcoding the sum, so the corner stays correct when either token moves.
Border Style
Borders across all form components and containers should use a highly restricted set of tokens.
The 2-Step Rule: Limit border widths to at most two options (e.g., 1px and 4px, or 1px and 8px). Do not use an incremental scale like 1px, 2px, 3px, 4px.... A limited choice makes the hierarchy clear and the product feel intentional.
--border-width-thin: 1px; /* Default for inputs, cards, dividers */
--border-width-thick: 4px; /* Featured items, bold accents, active indicators */
--border-color: var(--color-border);
--border-color-focus: var(--color-primary);
--border-color-error: var(--color-error);
An input border and a select border are identical at rest. Focus state uses --border-color-focus everywhere. Error state uses --border-color-error everywhere.
Spacing and Height
Components at the same visual scale share height and internal padding.
Heights and text padding are in rem so a control still contains its label when the user enlarges text; borders and shadows stay in px. See [[sizing-units]].
A button and an input placed next to each other must be the same height. This is not cosmetic — mismatched heights break form layouts and signal disorder.
Set the height, do not derive it. A control sized only by padding has a height of line-height + padding + border, so two controls in one row drift apart whenever any of those three differ:
an outlined variant beside a borderless one is 2px taller,
a fluid or clamped font size changes the line box at some viewports and not others,
an item whose content is an avatar or icon rather than text has a different intrinsic height.
Give every control on a line the same explicit height and centre its content. Set it as min-height, so the shape holds at rest and yields rather than clips when the content is larger than you planned.
Argue this at the box model, not at the pixels. Whether a border is part of the height is box-sizing, and naming that is what ends the discussion — a row debated in nudges never closes, because every participant is describing a different symptom of the same declaration. Under content-box the border is added to the height, so an outlined control is 2px taller than its ghost sibling and, once centred, sits 1px off it. Under box-sizing: border-box — the default in Tailwind and most resets — the border is absorbed into the height, so the boxes match and a variant can gain or lose its border without moving anything.
border-box moves the problem inward rather than removing it: the border now eats 1px of interior at the top and bottom, so a control whose content is not centred, or whose remaining interior is an odd number of pixels, still loses a rounding to one side. That is the 1px nobody can grep for. The fix is to keep the border in the box at all times — border: 1px solid transparent on the borderless variants — so the interior is identical whether or not a border is drawn.
A derived height is also a defect you cannot search for. An explicit height is one token you can grep and diff. A derived one is an emergent property of three separate declarations, so a row can be wrong in one control out of eight and no query finds it: in utility-class codebases the same padding appears in different orders (rounded-md px-3 py-2 and rounded-md border transition-colors px-3 py-2), and a find-and-replace fixes some of them and silently skips the rest. Each miss is 2px, invisible on its own, and the reason the row still looks broken after you "fixed" it.
The One Way: One Class Owns the Row
Repeating the same values across siblings is how the row drifts, because every later edit has to find every copy. Declare them once instead. Every item in a row uses one shared class; that class owns the box, the content slot, and every state; an instance may set colour and nothing else.
.control {
box-sizing: border-box;
height: var(--control-h); /* set, never derived */
display: inline-flex;
align-items: center;
gap: var(--control-gap);
padding-inline: var(--control-px); /* including its responsive steps */
border: 1px solid transparent; /* borderless variants keep the border, transparent */
border-radius: var(--radius-base);
font-size: var(--control-font); /* declared here, not on the label inside */
cursor: pointer; /* browsers give `button` cursor: default */
}
/* Opt-in, not `.control > *`: a descendant selector reaches icons that set
their own dimensions and stretches them. One content slot for every child
that carries content: avatar, label, icon, count. Equal boxes do
not make an equal row, and a control holding an avatar next to one holding a
text line looks uneven precisely when both measure identical. Standardising
the slot is not drawing everything at one size: inside a 20px slot an icon
can be 16px and a chevron 12px and the row still reads level. */
.control-slot { height: var(--control-slot); display: inline-flex; align-items: center; }
.control:hover { /* one definition for the whole family */ }
Three rules keep it true:
Nothing in the row sets height, padding, font size, radius or a state on itself. If one control needs something the class lacks, add a variant to the class.
A variant may change colour and nothing else. The moment a variant touches the box, it is a second class pretending to be one.
States are edited on the class, never on one instance. This is the regression that actually happens: the boxes are built correct, then months later one sibling gets a new hover, a new focus ring or a new transition and the row splits. A single row of eight controls has one box edit and dozens of state edits over its life, so the state rule is the one that pays.
An icon's mass is not centred in its box
align-items: center centres the icon's box. The drawn shape inside that box usually is not centred in it, so an icon that measures level reads low or high beside its label. A star loads its head; a download arrow loads its base. Give each icon its own offset, and do not share one nudge across a set: the correction differs per shape, and a shared value necessarily overshoots one icon and undershoots another. Two commits pushing the same row in opposite directions is the signature of a shared constant, not of one of them being wrong.
Measure it instead of nudging until it looks right. Rasterise the glyph, take the centroid of the alpha channel, and compare it to the centre of the box it will be centred in:
from PIL import Image
a = Image.open("icon.png").convert("RGBA").split()[3] # alpha = ink coverage
px, w, h = a.load(), *a.size
tot = sum(px[x, y] for y in range(h) for x in range(w))
cy = sum(px[x, y] * y for y in range(h) for x in range(w)) / tot
offset_px = (h / 2 - cy) / (h / RENDER_PX) # negative: lift the icon
Bake the result into the markup as a per-icon offset, name the measurement in a comment, and re-measure when an icon is swapped. Two things that sound right and are not:
Blurring first rarely changes the answer. Approximating how the eye integrates mass is a reasonable instinct, but on compact solid glyphs a Gaussian blur moves the centroid by hundredths of a pixel. It earns its place only on shapes with thin extensions, which weight a bounding box without weighting the eye.
Colour does not move the centroid. On a single-colour glyph, luminance scales every pixel by the same constant and the centre of mass is unchanged. Colour changes how heavy the icon reads next to the text, which is a size and weight decision, not an alignment one.
The same reasoning applies to a lone letter used as a mark, where the offset follows the letterform's mass and legitimately differs in sign between two letters.
The cursor belongs to the class too.<button> renders with cursor: default in every browser, and a framework reset does not necessarily fix it: Tailwind v4's preflight does not. The cursor is the cheapest affordance a pointer user gets and the one that reads before any hover colour arrives, so a control that looks clickable and keeps the arrow reads as inert. It survives review precisely because the hover state usually is implemented and only the cursor is wrong. Verify rather than assume, since preflight contents change between majors: grep -n "cursor" node_modules/tailwindcss/preflight.css. An element made interactive without a native tag (<div role="button">) needs the cursor, a focus style and key handling; the cursor alone is the shallowest part of that.
When those controls are joined edge to edge, the family radius applies to the group, not to each child: the outer corners keep --radius-base and the inner joints go square, so the bar reads as one control. Collapse the shared borders to a single line too — two abutting 1px borders render as a 2px seam that no other member of the family has.
The same logic runs vertically when a control opens. A select, multiselect, date picker or combobox and the panel it drops are one seam: the trigger keeps the family radius on its top corners and squares off the bottom, the panel squares its top and carries the family radius at the bottom, so the open control reads as one taller shape rather than two stacked ones. Whatever sits at the seam — a hairline divider, a small caret notch, nothing at all — does not change the corners. The panel's border is the trigger's border continuing, at the same width and colour, and the trigger's height stays --control-h while it is open.
For a group that wraps several controls in one shared surface (a balance beside an avatar, a segmented control, an input with an attached button) pin the height on the wrapper and set it on the children too. Stretching alone is a layout side effect that a later align-items change or an absolutely positioned child quietly removes.
Introducing the class is the dangerous step, and it fails in two specific ways. Both are silent in review and obvious on screen:
It must lose to the utilities it now sits beside. In a utility-first codebase every instance still carries colour classes, and a plain stylesheet loaded after the framework outranks them. Put the class in the framework's component layer (@layer components, or the equivalent @layer ordering), and write longhands, never shorthands: one border: 1px solid transparent repaints every button's border colour back to transparent, because the shorthand resets border-color that a utility had set. The same trap applies to background, padding, font and transition.
It must consume the existing tokens, not restate their values. Copying 8px out of the old markup as a literal forks the token: the row is correct today and stops following the design system the next time the token moves. Reference var(--radius-md), var(--control-h) and so on, and if the value you need has no token, add one.
After introducing it, re-measure. The class can be correct and still land wrong, and the measurement takes seconds.
Measure; do not reason. A wrong box, a right box with the wrong content mass, and a right row spoiled by a surface treatment all look like "the heights are off", so guessing between them fixes the wrong thing. Render the row against the app's real compiled stylesheet in a headless browser, at two viewport widths, and read every control:
Equal height proves the boxes match; equal top proves they share a baseline; differing fontSize or padLeft is the paradigm mismatch, as a number rather than an argument.
Check every shell that builds the row. The same toolbar is usually assembled in several places, and the one on screen may not be the one you opened. A fix that changes nothing visible means you edited the wrong file, not that the fix was wrong.
Grouping Non-Buttons Beside a Button Row
A toolbar often has to carry items that are not buttons: a balance, a status, an avatar, a count. Dropped loose into a row of buttons they read as broken buttons. The fix is common region (see [[gestalt-ui-organisation]]) — one surface that says "these belong together and are not the same thing as those" — and the surface, not the item, carries the grouping.
Ways to draw that region, quietest first:
Treatment
Reads as
Use when
Background tint (2–8% neutral)
A resting surface
Default. Quiet enough to sit beside outlined buttons without competing
Border
A container
The row already has borderless buttons, so a border still distinguishes
Inner shadow
Recessed, a well
The group is an input-like or display region rather than a set of actions
Gradient
Raised and physical
Rarely. It re-adds the button affordance you were trying to remove
Whichever you choose:
Match the buttons' border-radius exactly. A different radius beside them reads as a foreign element, not a sibling.
Match the height. Same rule as above; the group is one control's worth of vertical space.
Pick one treatment. Tint plus border plus inner shadow is three ways of saying the same thing, and at toolbar scale the element cannot absorb the weight (see Small Component Restraint).
Keep the children plain. No fills, no borders of their own — the region already grouped them. Their only state is hover, slightly stronger than the resting surface.
Do not let a child's hover redraw the group. Give the wrapper its own border and divider colours, static ones. If members reuse the standalone button style, hovering a single segment restyles the whole unit's outline, and the group appears to twitch.
One icon per group, not one per segment. Three segments each carrying the same download icon is the icon repeated three times, not three labelled choices. Show it only where the label cannot fit, e.g. at narrow widths.
Direction carries the meaning. Raised and recessed are the same two effects pointed opposite ways, and they make opposite promises — pressable versus readable. A recessed surface needs all three parts: the gradient running dark to light downward, the inset shadow on the top edge, and a light hairline on the bottom edge to close the well. Skip the last and the top shadow reads as grime rather than depth. Flip the first two and you have rebuilt the lit-from-above button you were trying not to be.
Tune both themes separately — the same alpha values that read as a subtle well on light read as flat or as a smear on dark.
Shadow
Interactive components use a consistent shadow logic:
At rest: no shadow, or --shadow-xs for floating components (select dropdown trigger)
On focus: focus ring via outline, not box-shadow (unless using box-shadow as the focus ring consistently)
Elevated (dropdowns, popovers opening from components): --shadow-md
Colour Logic
The same colour roles apply uniformly across all components:
State
Colour token
Rest border
--color-border
Focus border / ring
--color-primary
Error border
--color-error
Disabled
--color-text-secondary at reduced opacity
Selected / active fill
--color-primary
Hover background
--color-primary at 8–12% opacity
One Interaction Language
The colour table above defines what each state looks like; this rule governs how many interaction patterns a product is allowed to have. Pick one and reuse it — don't run 3–4 different hover/active/interaction patterns within the same product.
The user already knows what site they're on. Variety between products is expected; variety within one is taxing — every new pattern is another thing to learn mid-task. So converge:
One hover response. If interactive elements lift on hover, they all lift; if they shift background tint, they all shift tint. Don't mix lift, underline, colour-swap, and scale across sibling components.
One active/pressed response, one focus ring, one selected treatment — applied identically everywhere (this is why the colour roles above are shared tokens, not per-component choices).
One motion signature — the same easing and duration for the same kind of transition, so hovers, reveals, and toggles feel like one hand made them (see micro-interactions).
A tight, repeated interaction vocabulary is what makes a product feel learnable: the user learns the pattern once and trusts it everywhere.
Component Family Members
Component
Shares radius
Shares height
Shares border
Shares colour logic
Button
✓
✓
— (filled)
✓
Input / textarea
✓
✓
✓
✓
Select
✓
✓
✓
✓
Checkbox
--radius-sm
—
✓
✓
Radio
--radius-full
—
✓
✓
Toggle / switch
--radius-full
✓
—
✓
Pill / tag
--radius-full
✓
✓ optional
✓
Badge
--radius-sm or --radius-full
—
—
✓
Date picker / calendar
--radius-base
✓
✓
✓
Slider
--radius-full (track + thumb)
—
—
✓
Segmented control / button bar
✓ outer corners only
✓
✓ shared
✓
Search input
✓
✓
✓
✓
Combobox
✓
✓
✓
✓
Family Resemblance, Distinct Roles
Shared DNA makes components look related — it must not make them look interchangeable. The riskiest pairs are the ones that share the most: a pill-shaped button next to a pill-shaped badge, a bordered button next to a bordered input. When they blur, users click badges that do nothing and skip buttons that looked like labels.
The rule: role must be readable before interaction. From appearance alone, the user can tell what is clickable, what is editable, and what is read-only.
Button — clickable: solid fill or a firm border, cursor: pointer, a hover response.
Badge / tag — read-only: muted fill, smaller type, no hover response and no pointer cursor, ever — those two signals are reserved for interactive elements and are exactly what separates a badge from a button of the same shape.
Input — editable: border with an empty interior, placeholder, text cursor.
An input and a button side by side are the pair that fails most often. Both are a rounded box of the same height with a 1px border, and once the button is a ghost or outline variant the two are the same drawing. Keep the height identical — that is family — and put the difference in the interior: the input stays empty with a placeholder and a text cursor, the button carries a fill or a firm border, a label set in the button weight, and a pointer. If the only thing telling them apart is that one has text in it, the form is one placeholder away from ambiguity.
Distinguish through at least two visual channels (fill + size, border + cursor) — never by colour alone, and never by height. Squint test: with labels unreadable, can you still sort the buttons from the badges from the inputs? If not, the family has collapsed into one component.
Semantic Chip Components
Generic Badge components lead to misuse — the same component ends up used for statuses, code tokens, keyboard shortcuts, and categorical labels, with style overrides scattered across the codebase.
The pattern: one component per meaning, not one component with many variants.
A product's inline labels typically fall into a small set of distinct meanings. Define a component for each one. Common examples:
Component
Meaning
Shape
Tag
Categorical label, status, filter
Pill (rounded-full)
Code
Inline literal, path, key
<code>, mono, tight radius
Kbd
Keyboard shortcut
<kbd>, mono, tight radius
Metric
Measured value (1.2s, 42px)
Mono, tight radius
Add product-specific types as needed (e.g. Flag for CLI products, Token for API products). Each new type gets its own component — not a new variant prop on an existing one.
Each component encodes exactly one meaning. Appearance follows from it — callers never pass colour or shape props.
Sizing: use em-relative padding so a chip renders at the right size for whatever text context it sits in (heading, body, caption) without per-context overrides.
Colour: keep per-semantic colours in CSS classes (chip-tag, chip-code, etc.) in one file. Do not inline colour props. This keeps light/dark mode in one place and lets you audit the full chip palette at a glance.
Back-compat: if existing call sites use a generic Badge, re-export the most common semantic variant as the default so old imports keep working without a migration.
Alignment on a Line
When a line mixes element types — label, badge, status dot, value, icon — they must read as one aligned row, not a jumble of differently-sized pieces.
Same type size on the line. Text next to a badge or chip shares the surrounding typeface size; a badge must not silently shrink or enlarge its row. Use leading-none and align-middle (as in the chip BASE above) so every element sits on a shared centre line.
Centre status indicators. A traffic-light dot (red/amber/green) is vertically centred against the text it annotates — aligned to the label's cap-height centre, not the baseline.
A badge beside text sits on the text. Align it to the baseline or centre it optically on the x-height. Grow badge padding symmetrically so it stays centred against its neighbour, not only inside itself.
One optical centre line. If badges, text, and icons jump up and down, the row reads as broken even when each piece is fine alone.
A count badge stays a circle
Size a count badge with border-box, a height and a min-width of the same value, and horizontal padding only, so it widens for a second digit. Do not put a fixed width on a generic pill badge that already carries side padding and a border: content-box adds both, and a 20px circle renders as a 42px pill that reads as a label.
Small Component Restraint
The smaller the component, the less it can carry. Restraint that looks plain at large sizes is what keeps small components legible.
Avoid multi-border / boxed-in containers. Don't stack bordered layers (a bordered chip inside a bordered cell inside a bordered card) — a small element can't absorb the weight. Prefer one border or none; use fill or spacing instead. The small-scale companion to the 2-Step border rule.
At most one icon. Two or three icons in a pill, badge, or row create clutter and ambiguity about which is actionable. If you need more, the component has outgrown its size — promote it to a larger pattern.
Multi-card-container design → pivot. Card-in-card-in-card nesting solves grouping with boxes instead of spacing and hierarchy. Flatten it, group with whitespace and headings (see [[gestalt-ui-organisation]]), and keep the card metaphor for the outermost meaningful container only.
If the brand uses gradients, apply them consistently:
A gradient on a primary button should use the same gradient angle and stops as gradient usage elsewhere in the product
Hover state: slightly shift the gradient lightness, not the hue
Do not use gradients on some button variants and flat colour on others — pick one approach per variant and apply it universally
Reuse the Product's Own Marks
A new view that needs marks for states (same, changed, added, removed) searches the product for the marks it already uses and reuses them exactly. A second vocabulary for the same idea teaches the reader two languages and contradicts meanings fixed elsewhere.
Review Checklist
Do buttons and inputs on the same form share the same height?
Is that height set explicitly rather than left to add up from padding, line-height and border?
Do all controls in a row come from one shared class, rather than repeating the same values?
Does that class sit in the component layer and use longhand properties, so instance utilities still win?
Does it reference the existing radius, height and spacing tokens rather than copying their values?
Does every child of a control (avatar, label, icon) sit in the same content slot height?
Is every state defined on that class, so a new hover or focus ring cannot land on one sibling only?
Were the heights measured with the real stylesheet at more than one viewport, rather than reasoned about?
Do all bordered components use at most two border-width options (e.g., 1px and 4px)?
Does focus state look identical across all focusable components?
Does error state look identical across all components that can have errors?
Is there a single interaction language — one hover response, one active/pressed response, one focus ring, one motion signature — reused across the product, rather than 3–4 competing patterns?
Are all radius values derived from the same base token — not set independently per component?
Do pills and tags use --radius-full consistently?
Can a button, badge, and input be told apart from appearance alone (squint test) — with non-interactive elements carrying no hover response or pointer cursor?
Is gradient usage (if any) consistent across all button variants?
Before building a new component, was the existing library (and codebase) checked for one that fits as-is, or a one-off that could be generalised with a small change, rather than cloning?
Could a new component be added to the library using only existing tokens?
Are inline labels, statuses, code tokens, keyboard hints, and metrics separate components — not variants of a generic Badge?
Do chip/badge components use em-relative sizing so they scale with their text context?
Is chip colour defined in CSS classes (not inline props) so light/dark lives in one place?
On rows mixing text, badges, and icons, does everything share one type size and centre line?
Are status/traffic-light indicators vertically centred against their label?
Do small components avoid stacked/nested borders (boxed-in look)?
Do small components carry at most one icon?
Has card-in-card-in-card nesting been flattened in favour of spacing and hierarchy?
Does a new view reuse the state marks and icons the product already uses, with none invented locally?
