Design Tokens

Design Tokens

Announce at start: "I'm using the humane:design-tokens skill to set up, resolve, and export this token set."

Manage DTCG 2025.10 design tokens with a dependency-free Python core. v1 covers the deterministic spine: scaffold, validate, merge (global base + project override), resolve aliases, and export CSS.

Standard vs convention

• Standard (DTCG): *.tokens.json, $value/$type, whole-value {alias} references.
• Skill convention (NOT DTCG): global-base / project-override layering via merge, and theme-as-override-file. These are labelled in code; do not present them as standard.

v1 scope

Supported $type: color (string values), dimension, duration, fontFamily, fontWeight, number, typography, shadow. Outputs: CSS custom properties, a Google-Labs DESIGN.md (alpha), a standalone HTML preview, and generation prompts (gpt-image-2 / nano-banana CLI lines + a /tufte-report theme) via the prompt door. CSS import covers color/dimension/duration/fontFamily/number; composite values (box-shadow, gradients, multi-part typography) are skipped and reported. Not in v1: JSON Pointer $ref, $root, structured color objects, name-restriction enforcement, Style Dictionary, Figma/Pencil importers, share bundles, skillify (see the phased spec).

Commands

Run via scripts/tokens <command> (or PYTHONPATH=scripts python3 -m dtokens.cli):

Contrast

DTCG stores colors, not relationships — nothing in the standard says which token is text and which is the surface behind it. contrast adds that layer as a SKILL CONVENTION and makes it executable, so a token set cannot compile with an unreadable role pair.

Two scales, both reported:

--standard both (the default) requires clearing both. They genuinely disagree: #747474 on white is 4.67:1 (passes WCAG AA) but Lc 72.5 (fails APCA). APCA is the better predictor of perceived readability; WCAG is the one auditors ask for. Reporting both, and letting the caller pick which gates, is the honest split.

Fixes move lightness only. The suggestion walks OKLCH L away from the background until the pair clears, keeping C and H — so the brand hue survives the fix. When no lightness on that axis clears the bar, no fix is offered: chroma or the background has to move, and that is a design decision, not a mechanical one.

Unparseable colors are reported, never failed. var(--x), currentColor, a gradient, or any value carrying alpha (contrast depends on what is behind it) is listed as not measured. A verification gap is not a finding.

Palette ramp steps are skipped. A trailing numeric step (ink-950, amber-500) marks a swatch, not a role assignment — ink-950 is a color that happens to contain "ink"; text is the token that says where ink goes. Pairing ramp steps produced confident nonsense on real sets, so roles are read from semantic names only.

Two names resolving to the same value fail. A declared foreground and background that carry one color is invisible text — Lc 0 at 1.0:1 — and the worst thing this command could do is call it a benign alias and exit 0. It is reported as the most severe failure there is, with no proposed fix: no lightness move on the foreground repairs a pair that is wrong in itself. Two unrelated neutrals sharing a hex are still listed, but they only reach the gate when the set declares them as a pair.

Declaring the pairs that actually meet

Name inference cannot know intent. Our own set defines paper-50 — a warm paper surface for printed / risograph contexts — which reads as a background by name but never sits behind screen text. Declare the real pairs to fix this permanently (SKILL CONVENTION, at the token-file root):

"$extensions": {
  "community.design-tokens.contrast": {
    "pairs": [["text", "background"], ["muted", "background"], ["on-primary", "primary"]],
    "exclude": ["surface", "accent"]
  }
}

• pairs — when present, these are measured and nothing else. Each entry is [foreground, background], optionally [foreground, background, "body"|"non-body"|"graphic"]. Names resolve as a full path (color.text), a flat name (brand-primary), a bare final segment (primary), or a role. Where a short name is ambiguous across groups, a candidate in the same group wins — so color.brand.on-primary pairs with color.brand.primary, never with a color.chart.primary that merely shares the leaf.
• exclude — tokens never paired in either position; applies whether pairs are declared or inferred.
• pairs: [] measures nothing. An empty list is a declaration that no pair meets, not an absent declaration — it does not fall back to inference.
• Absent entirely, pairs are inferred from roles, so existing sets keep working.
• A name that resolves to no token is reported, never dropped. A typo in a declared pair prints under named in the contrast declaration but not found and exits non-zero. Silence there would leave the gate green over exactly the pair you asked it to check. The same applies to an unknown level.

Levels. body (Lc 75 / 4.5:1) is text read in quantity; non-body (Lc 60 / 3:1) covers links, icons, badges, and large display text; graphic (Lc 45 / 3:1) is color that is never text — a fill, a rule, a chart mark. graphic must be declared, per pair or via --level graphic: inference never assigns it, because a token's name cannot tell you whether it is painted as type. An on-X token is always measured as body — it is ink by definition, whichever fill it names.

Gating

• validate <file> and use print failures as warning: lines — advisory, exit 0.
• validate <file> --strict promotes them to errors (exit 1), so CI can gate on them.
• contrast <file> exits 1 on any failure by itself.

Themes

A token file is one theme (the convention is one override file per theme, merged before use). So run contrast after the merge, once per theme — a palette verified only in light mode is a palette half-verified.

Brand-style extensions ($extensions)

SKILL CONVENTION: a $extensions["community.design-tokens.brand"] block at the token-file root (mood adjectives, imageryStyle/voice prose, subjects, avoid, negativePrompt) feeds the prompt door: mood/imageryStyle join the brand clause, avoid becomes DON'T lines, negativePrompt an "Avoid:" tail. See templates/brand-extensions.example.json (worked ai-design example). Fidelity tests (references/prompt-fidelity-notes.md): art-direction prose with hexes + color words beats both the bare comma-clause and strict constraint blocks (which are unreliable on Nano Banana); provider capabilities in references/providers.md.

The block is technically optional in DTCG, but leaving it out makes the art-direction contract silent: palette and type say nothing about illustration style, so every downstream generator (brand-illustrate, the prompt door) has to guess or ask. Treat it as a first-class part of setup, not an afterthought — author it during setup-edit, and validate/use will warn (never fail) when it or imageryStyle is missing.

Authoring the brand block (questionnaire)

After scaffolding a set (setup-edit), or when editing a set whose validate/​use output warns that the block is missing, walk the user through these questions one at a time. Use a structured multiple-choice question tool when the agent has one; fall back to a plain numbered question otherwise. Write the answers into $extensions["community.design-tokens.brand"].

Claude Code extras: use AskUserQuestion for the pick-one/pick-many steps (imageryStyle direction, mood adjectives) so the options render as chips. On other agents, list them as plain text and read the reply back.

1. imageryStyle (load-bearing — never skip). "What visual language should on-brand imagery use?" Offer concrete directions, let the user pick one or describe their own:
   • flat-geometric — flat vector, geometric shapes, dot-grid/stipple textures, code-native
   • technical-line — thin-line technical/blueprint drawing, engineering diagram feel
   • risograph — textured print, limited flat inks, grain, zine aesthetic
   • painterly — brush/ink or gouache, organic, hand-made
   • photographic — real photography, lighting/lens language
   • 3D-sculptural — rendered forms, material and depth
   • Other — the user's own prose (capture verbatim) Store as a short prose sentence (e.g. "flat vector illustration with dot-grid textures; no photorealism"), not just the keyword — fidelity tests show prose beats a bare tag.
2. mood — "Two to four adjectives for how the brand should feel." (e.g. precise, editorial, calm.) Store as a list.
3. avoid — "Anything specific this brand must never show?" Brand-specific negatives (stock-photo people, glossy 3D blobs, lens flares). Store as a list; these become DON'T lines. (The generic de-slop negatives are added downstream by brand-illustrate, so don't restate them here.)

Optionally capture voice, subjects, and a terse negativePrompt tail when the user offers them. Confirm existing values on edit rather than re-asking from scratch.

Serving previews (default)

Generated HTML is meant to be served, not opened from disk. Browsers treat file:// URLs as unique security origins, which breaks cross-origin web-font loads, fetch, and many extensions (Unsafe attempt to load URL … 'file:' URLs are treated as unique security origins). So use (and preview when it writes a file) start a tiny stdlib HTTP server on http://127.0.0.1 and open the result — by default when run interactively (a TTY). In scripts / CI (non-TTY) serving is skipped so nothing blocks; force it either way with --serve / --no-serve. serve <path> does the same for any existing file or directory. Dependency-free (http.server).

Extracting tokens from a live site

Tokens can be reverse-engineered from any shipping site, then saved as a template. The method doesn't matter — pick what the site allows:

1. Fetch the CSS (static sites): grab the linked stylesheet(s), then resolve the variable indirection to ground values. Modern design systems alias twice — e.g. Monaspace's --color-neon-primary: rgb(var(--color-neon-primary-rgb)) and --color-neon-primary-rgb: 245 184 165 → #F5B8A5. Base scales (--base-size-16: 1rem) give the spacing/radius steps.
2. Computed styles (JS-rendered sites): drive a real browser (/browser-mate) and read getComputedStyle(:root) plus key elements — yields ground-truth values no matter how they're authored.
3. import <css>: if the site exposes a flat :root block, pipe it straight through the importer (names preserved).
4. Hand-curate the extracted values into <name>.tokens.json with explicit role aliases (primary, text, background, …) so the prompt door and tufte map light up, and validate.

templates/monaspace.tokens.json is the worked result for monaspace.githubnext.com — its five-font superfamily (Neon/Argon/Xenon/Radon/Krypton) as accent colours over the GitHub dark canvas (#0D1117), on the 4/8/16/24 base scale. Scaffold from it with setup-edit my.tokens.json --from templates/monaspace.tokens.json.

Prompt door (tokens → generation)

The spine ends at CSS/DESIGN.md; the prompt door carries the brand onward into image and report generation so it never needs hand-translating. All of this is skill convention, not DTCG — colour roles, the curated preset picks, and the tufte variable map are generation aids layered on the standard.

• Brand summary (brand_summary.py): distils resolved tokens into palette (with roles inferred from token names: primary, text, background, accent, success, warning, danger, muted), fonts, type specimens, and a shape word from the largest corner radius (sharp/soft/rounded/pill).
• gpt-image-2: emits CLI lines across that tool's unique presets (editorial, bauhaus, isometric, poster) so one brand yields distinct moods; the brand's exact hex/fonts/shape are baked into each subject.
• nano-banana: steers to its edge — accurate in-image text (--model pro) and reference-image anchoring — over the shared presets. (It has no presets unique to itself; its set is a subset of gpt-image-2's shared eight.)
• tufte: emits a :root block mapping brand roles onto /tufte-report's own variables (--ink, --bg, --spark-primary/secondary/tertiary, --status-red/amber/green, --accent). Roles with no matching token fall back to tufte-report's defaults, labelled inline. (/tufte-report consumes a theme, not a DESIGN.md.)

DESIGN.md output

use and design-md emit a DESIGN.md — the agent-facing format read by Claude Code, Cursor, v0, Lovable, Stitch. It is complementary to DTCG: DTCG .tokens.json is the rigorous source of truth; DESIGN.md is the prose+tokens artifact agents apply. Our resolved tokens map to its frontmatter as: color → colors, typography → typography, dimension under space* → spacing, dimension under radius/rounded* → rounded. Names are flattened (drop the top group, dots → -). Types without a DESIGN.md home (duration, shadow, number, fontFamily, fontWeight standalone) are noted in the Overview, not the frontmatter. This name/bucket mapping is a skill convention over the DESIGN.md alpha schema.

Generated, not hand-edited. DESIGN.md is a compiled render of the token set — the same one-way pipeline as CSS. It carries provenance frontmatter (generator: "design-tokens", source:, regenerate:) plus a do-not-edit line in the Overview naming the source file and the exact regeneration command. setup-edit writes it next to the token file (its canonical home per the storage convention); use writes it into the output dir; design-md prints or -o-writes it. Stale-overwrite guard: before overwriting, an existing DESIGN.md that lacks the generator: "design-tokens" marker is treated as hand-written/foreign — the tool warns and leaves it untouched rather than clobber it; regenerate explicitly with design-md <src> -o DESIGN.md.

Brand direction (default output). Whenever the $extensions brand block carries mood / imageryStyle / subjects / avoid, a ## Brand direction section renders in the default DESIGN.md (not gated behind --rich) — so the art-direction contract travels with the compiled artifact. The frontmatter schema is otherwise unchanged for Labs consumers.

Rich mode (opt-in, non-standard). --rich sources extra sections from optional keys in $extensions["community.design-tokens.brand"]: essence (prose), components [{name, role, spec}], animations [{name, role, spec}] (spec may embed code fences — motion recipes render under "Animation Recipes"), dos/donts [str], surfaces [{level, name, value, purpose}], elevation {label: note}, imagery (prose, falls back to imageryStyle), layout (prose), similarBrands [{name, note}|str]. The frontmatter stays standard; the body gains a labelled skill-convention block, so the file is no longer a plain Labs alpha document — hence the confirmation prompt.

Storage convention

Token files are canonical source — keep them visible and committed, never in a hidden dotdir (a leading dot reads as ignorable tool state; Style Dictionary uses tokens/, DESIGN.md lives at repo root, DTCG mandates the extension but no path).

• Global sets: ~/design-tokens/<set>/base.tokens.json (a shared, user-level location outside any project — put your agent's config dir here if it has one)
• Project, single set: <project>/design.tokens.json + <project>/DESIGN.md at root
• Project, multi-scope: <project>/tokens/base.tokens.json + <project>/tokens/<name>.tokens.json
• Multiple themes (light/dark): keep one override file per theme and merge it before use.
• Reserve a dotdir, if any, only for generated output (tokens.css, preview.html).

Tests

cd design-tokens && PYTHONPATH=scripts python3 -m pytest tests/ -v