Design System

Document Design System

The visual layer for HTML documents and slide decks: a token file in, a single self-contained CSS bundle out, with every color pairing checked against WCAG before it ships. This is the theming layer for documents — type scales, reading measure, print roles, light/dark surfaces. It is not a product UI component library; there are no buttons, form states, or component variants here.

When to use this skill

• Theming a report or whitepaper that must render as one self-contained HTML file
• Building a deck theme where type must stay legible at projection distance
• Auditing an existing palette before a public or regulated publication
• Diagnosing dark-mode drift — text that reads fine in light mode and fails in dark
• Standardizing across documents so a series of reports looks like one series
• Answering "does this pass AA?" with a number instead of an opinion

Inputs the skill expects

• A design-token JSON file (or the intent to generate one from assets/sample_tokens.json)
• Brand colors, if any exist — a hex value and where it came from
• Typeface choices plus their fallback stacks (self-contained HTML cannot fetch web fonts)
• The conformance target: AA (default) or AAA
• Which artifacts consume the theme: document, deck, or both
• Whether the output will be printed or exported to PDF

Clarify First

Before generating a theme, confirm these inputs. If any is unknown or vague, ASK — do not assume:

• Conformance target (AA vs AAA) — why it changes the output: AAA at 7:1 rules out most saturated mid-tones, so the accent and warning colors must be picked differently from the start
• Print / PDF export required — why it changes the output: print needs a forced light role set and page-break rules; retrofitting them means re-deriving every dark value
• Document or deck — why it changes the output: the type ratio differs (1.25 vs 1.333) and decks need larger minimum sizes for projection
• Existing brand colors — why it changes the output: a fixed brand hex constrains the accent ramp and may force the underline-links decision

Stop rule: ask only the 2-3 that most change the output. If the user says "just draft it," proceed and list your assumptions at the top of the artifact.

Workflows

Workflow 1 — Generate a themed CSS bundle from tokens

1. Copy assets/sample_tokens.json and edit the palette, roles, and scales.
2. Lint the token file first — a malformed ramp produces a valid-looking bundle with inverted colors.
3. Compile the bundle and inline the result into the target HTML document.

python3 markdown-html/design-system/scripts/token_linter.py \
  --input markdown-html/design-system/assets/sample_tokens.json --format text

python3 markdown-html/design-system/scripts/theme_builder.py \
  --input markdown-html/design-system/assets/sample_tokens.json \
  --out build/theme.css --format text

Workflow 2 — Gate a palette on WCAG contrast

1. Declare every real foreground/background combination in the pairings block, each with its usage class.
2. Run the AA gate; it exits non-zero on any failure, so it drops into CI directly.
3. Run --all-pairs to find combinations nobody declared but a stylesheet will eventually produce, and --level AAA --no-gate as an aspirational report.

python3 markdown-html/design-system/scripts/contrast_validator.py \
  --input markdown-html/design-system/assets/sample_tokens.json \
  --level AA --format text

python3 markdown-html/design-system/scripts/contrast_validator.py \
  --input markdown-html/design-system/assets/sample_tokens.json \
  --level AAA --no-gate --format json

Workflow 3 — Audit and repair an inherited theme

1. Run the linter to surface structural rot: non-monotonic ramps, roles present in one mode only, literal hex values that will not respond to theming.
2. Run the validator with --all-pairs to get the full contrast matrix.
3. Fix in this order — ramp order first, then mode parity, then contrast. Ramp and parity errors invalidate the contrast numbers, so fixing contrast first wastes the work.

assets/sample_tokens_legacy.json is a deliberately damaged theme carrying all four failure classes, so this workflow demonstrates real repair rather than a clean run. Expect findings from both commands: the linter reports 4 errors and 5 warnings and exits 2; the validator is pinned to report-only with --no-gate and finds 13 failing pairings.

python3 markdown-html/design-system/scripts/token_linter.py \
  --input markdown-html/design-system/assets/sample_tokens_legacy.json \
  --max-severity warning --format json

python3 markdown-html/design-system/scripts/contrast_validator.py \
  --input markdown-html/design-system/assets/sample_tokens_legacy.json \
  --all-pairs --no-gate --format text

Decision frameworks

Type ratio selection

Use 1.25 for documents and 1.333 for decks. A document needs 7-8 usable steps from caption to H1; at 1.5 the top of that range is 8.5x the base, which no report can place on a page. Escape hatch: a single-page poster or title card can use 1.5 because it has one heading and no hierarchy to preserve.

Contrast usage classes

Classify honestly. A border that is the only thing separating two data regions is ui, not decor. The 1.5:1 decor floor is not a WCAG number — it is the point below which a rule stops being visible on a mid-quality screen, so it fails at its decorative job too.

Contrast targets beyond the minimum

[RECOMMENDED] Do not use pure black on pure white. 21:1 is the maximum and it is worse than ~16:1 for extended reading — glyph edges bleed on bright displays, and readers with astigmatism report the most discomfort at that pairing.

Dark-mode role derivation

Moving an accent only one step is the most common dark-mode bug in this domain: accent.600 scores 5.9:1 on white and 3.4:1 on near-black, so it passes the light gate and fails the dark one.

Anti-Patterns

The mode-in-the-name role

Mistake: Defining light-text and dark-text as two separate roles instead of one text role with two mode values. Why it happens: It mirrors how the designer thinks — two comps, two palettes — and each role reads unambiguously in isolation. Instead: Keep one semantic role and let roles.light and roles.dark supply the values. The mode belongs in the mode map, not the name. The linter's parity check enforces this by failing any role defined in only one mode.

The literal hex escape hatch

Mistake: Hard-coding one color directly in a component rule because no existing role quite fits. Why it happens: Adding a role feels like ceremony for a one-off, and the deadline is real. Instead: Add the role or reuse the nearest one. That hard-coded value will not respond to theming and surfaces months later as the single element that stays dark in light mode. The linter flags it as ROLE_LITERAL specifically because it is always cheaper to fix on the day it is written.

Untested dark mode

Mistake: Authoring the light theme carefully, mechanically inverting for dark, and never rendering the result. Why it happens: Dark mode looks like a mechanical transform, and the CSS compiles either way. Instead: Run the contrast validator across both modes — it checks every mode in the roles map for exactly this reason — and then actually open the document in dark mode. The failures cluster in the mode nobody looked at, especially in saturated accent and warning colors.

Trusting a green contrast run

Mistake: Treating a passing validator as proof the theme is accessible. Why it happens: The tool gives a number and the number is above the threshold, which feels conclusive. Instead: Read the list of what pairing math cannot see: text over images, semi-transparent overlays, hover and focus states nobody declared, two chart series that both pass against the background but not against each other, and red/green pairs at equal luminance that pass every ratio test and vanish for a deuteranopic reader. Contrast ratio is a luminance metric and is hue-blind by construction.

The 40-role sprawl

Mistake: Adding a role per component — table-header-bg, figure-caption-color, toc-link-hover. Why it happens: Each addition is locally reasonable; no single one looks like a mistake. Instead: Treat roles as a vocabulary, not a mapping table. Eleven roles cover a document; twenty is a smell; forty means the semantic and component layers have merged and the theme can no longer be re-skinned.

Files

All scripts share one exit-code contract: 0 clean, 2 gate failed (findings at or above the threshold), 1 the tool itself errored. A CI job can therefore tell a real defect from a broken invocation.