Design System — Onboarding + Shared Brand Tokens¶
The design-system skill is the shared brand owner for the markdown-html plugin. Run its onboarding once. Every converter (md-document, md-review, md-slides) reads the resulting config via config_loader.py and applies the same 12 CSS custom properties to its output. Without this, conversions render with placeholder defaults — technically functional but unbranded.
This skill ships exactly three Python tools:
onboard.py— interactive (or--defaults/--set/--show/--reset) wizard.config_loader.py— importable customization loader with project > global > defaults precedence andMARKDOWN_HTML_NO_CONFIG=1bypass.brand_palette_validator.py— WCAG-AA contrast checker + HSL palette deriver.
All three are stdlib-only and contain no LLM calls (deterministic per Path-B discipline).
When to invoke¶
| Symptom | Action |
|---|---|
| User says "convert this markdown to HTML" for the first time in this workspace | Run python3 markdown-html/skills/design-system/scripts/onboard.py |
~/.config/markdown-html/design-system.json doesn't exist OR setup_completed_at is null |
Refuse conversion, surface onboarding |
| User wants per-repo brand override | python3 .../onboard.py --scope project |
| User wants to change a single field non-interactively | python3 .../onboard.py --set brand.primary=#FF6B35 |
| User wants to reset and re-onboard | python3 .../onboard.py --reset then re-run |
| User wants zero-touch defaults (CI, ephemeral session) | python3 .../onboard.py --defaults |
| Headless / containerized run that should ignore saved config | MARKDOWN_HTML_NO_CONFIG=1 ... |
Onboarding question set (10 questions)¶
| # | Key | Choices / Validator | Default |
|---|---|---|---|
| 1 | default_output_dir |
path; os.access(parent, os.W_OK) |
./markdown-html-out/ |
| 2 | brand.primary |
HEX ^#?[0-9a-fA-F]{6}$ |
#0A1628 |
| 3 | brand.accent |
HEX or blank (auto-derive) | derive from primary |
| 4 | typography.heading_font |
Google Font name (12 safe defaults) | Inter |
| 5 | typography.body_font |
Google Font name | Inter |
| 6 | design_style |
editorial / technical / minimal / playful |
technical |
| 7 | code_theme |
light / dark / auto |
auto |
| 8 | toc.behavior |
sticky-sidebar / collapsible-top / inline / none |
sticky-sidebar |
| 9 | company_name |
string (may be empty) | "" |
| 10 | logo_url |
URL or empty (base64-embedded at render) | "" |
Hard rules¶
- WCAG AA body-text contrast must pass.
brand_palette_validator.validate()runs after every change. Body text on bg must reach 4.5:1; link on bg must reach 4.5:1. If either fails,onboard.pyrefuses to save (exit code 4) and tells the user to pick a darker primary, blankbrand.bg/brand.textto let derivation pick a safe pair, or overridebrand.textdirectly. Canon: WCAG 2.2 §1.4.3. - Output directory must be writable.
onboard.pywalks up the path to find an existing ancestor and checksos.W_OK. Empty or unwritable path → exit code 3. The orchestrator'soutput_path_resolver.pyhonors the same rule per-conversion. - Customization must change behavior, not sit as decoration. Every consumer (md-document, md-review, md-slides) must read the config and render differently when the user changes
design_style,brand.primary,code_theme, ortoc.behavior. Decorative-only fields fail the design discipline. - Precedence is fixed. Project > global > defaults. The deep-merge preserves nested keys (e.g. you can override
brand.primaryin a project config without losingtypography.heading_fontfrom global). - Bypass env exists for a reason.
MARKDOWN_HTML_NO_CONFIG=1is for headless CI, ephemeral test containers, and the autoresearch-style evaluator loops. Never set it silently for an interactive user.
Derived 12-token palette¶
Once the user's brand is captured, brand_palette_validator.derive_palette() produces 12 CSS custom properties stored under derived_palette in the same config file. Every converter inlines these into its <style> block.
| Token | Purpose | Derivation |
|---|---|---|
--md-bg |
Document background | Primary if dark, near-neutral if vibrant |
--md-surface |
Card / callout / blockquote background | Bg ± 4-6% luminance |
--md-border |
Hairline dividers, table borders | Bg ± 8-12% luminance |
--md-text |
Body text | Off-white on dark bg, near-black on light bg |
--md-text-muted |
Captions, metadata, footers | rgba(text, 0.68) |
--md-accent |
Primary CTA, callout headers, link emphasis | Primary if vibrant, hue-shifted lighter if dark |
--md-accent-soft |
Accent backgrounds, hover states | rgba(accent, 0.14) |
--md-code-bg |
Inline code, fenced block bg | Bg ± 4-5% luminance |
--md-link |
Hyperlinks | Iteratively walked to reach 4.5:1 contrast on bg |
--md-link-hover |
Hover state | Link ± 6-8% luminance |
--md-success |
OK / approved / passed | Green anchored, luminance-matched |
--md-warn |
Caution / nit / TODO | Amber anchored, luminance-matched |
Forcing-question library (Matt Pocock grill-with-docs pattern)¶
One question per turn, recommended answer, canon citation.
- What's your brand primary color? Recommended: a HEX you already use in your product or docs — not a stock blue. Canon: Aarron Walter, Designing for Emotion (color carries brand affect).
- Should accent be derived or set? Recommended: derive on first run (hue-shift + lighten produces a coherent companion); set explicitly only if your brand kit specifies one. Canon: Adobe Spectrum, Color Foundations.
- Editorial, technical, minimal, or playful? Recommended:
technicalfor engineering specs/reports,editorialfor long-read narratives,minimalfor sparse reference docs,playfulfor marketing/landing content. Canon: Ellen Lupton, Thinking with Type (style serves the rhetorical purpose). - Sticky-sidebar TOC, or inline? Recommended:
sticky-sidebarfor documents over 800 words,inlinefor short reads. Canon: Nielsen-Norman, Table of Contents Best Practices (2023). - Save to global or per-project? Recommended: global by default (consistent across your work); use
--scope projectonly when this repo has a different brand. Canon: research-ops onboarding pattern,research-ops/CLAUDE.md§8.
Customization in use (worked example)¶
# First-run onboarding (interactive, walks all 10 questions)
python3 markdown-html/skills/design-system/scripts/onboard.py
# Zero-touch defaults for CI / first-test
python3 .../onboard.py --defaults
# Change just the primary color and design style
python3 .../onboard.py --set brand.primary=#FF6B35 --set design_style=editorial
# Per-repo override
python3 .../onboard.py --scope project --set design_style=minimal
# Reset and re-onboard
python3 .../onboard.py --reset
python3 .../onboard.py
# Inspect the effective config (project > global > defaults)
python3 .../config_loader.py --show
python3 .../config_loader.py --status
# Bypass saved config (returns DEFAULTS only)
MARKDOWN_HTML_NO_CONFIG=1 python3 .../config_loader.py --show
# Spot-check WCAG contrast before committing to a brand
python3 .../brand_palette_validator.py --primary "#FF6B35" --accent "#00D4AA"
Assumptions¶
- User has at least one brand HEX they want consistent across their HTML conversions.
- User accepts a 1-2 minute one-time setup.
- User is OK with Google Fonts as the typography source (CDN, no local font hosting).
- WCAG 2.2 AA is the accessibility floor (4.5:1 body, 3:1 large/UI). AAA (7:1) is out of scope.
Non-goals¶
- Not a full design-token system (Style Dictionary, Theo). Twelve tokens, not a hundred.
- Not a custom-font hosting solution. Google Fonts only.
- Not a dark/light mode switcher in the converters.
code_theme: autohandles the prefers-color-scheme case for syntax highlighting; layout palette is single-mode per onboarding. - Not an accessibility audit suite (use axe-core / pa11y for that). We enforce contrast only.
- Does not transform existing CSS — the derived palette is injected into freshly generated HTML.
Distinct from¶
marketing/landing/skills/landing/scripts/brand_palette_validator.py— that script'sderive_palette()produces 8 tokens shaped for hero-page rendering (--navy,--teal,--card-bg,--card-border). This script produces 12 tokens shaped for document rendering (sticky surface, hairline border, code bg, link, link-hover, success, warn). Same WCAG + HSL math, different token taxonomy.research-ops/skills/clinical-research/scripts/onboard.py— same pattern (interactive +--defaults/--set/--show/--reset/--scope), different question set (clinical alpha/power/dropout vs. brand palette/typography/layout).
Output artifact¶
~/.config/markdown-html/design-system.json (global) or ./.markdown-html/design-system.json (project). JSON schema lives at assets/design_system_schema.json.
Anti-patterns (do not)¶
- ❌ Skip onboarding and run a converter with placeholder defaults — output looks unbranded.
- ❌ Pick a vibrant brand primary as
brand.bgdirectly (low text contrast). Use it as accent instead. - ❌ Set
MARKDOWN_HTML_NO_CONFIG=1silently for an interactive user — they'll wonder why their tokens disappeared. - ❌ Encode brand semantics in
derived_paletteoutside the 12-token taxonomy. Add a new token only with a deliberate name + purpose + derivation rule.
References¶
- WCAG 2.2 — §1.4.3 (contrast), §1.4.4 (resize), §1.4.11 (non-text contrast)
- Aarron Walter — Designing for Emotion (A Book Apart)
- Ellen Lupton — Thinking with Type
- Adobe Spectrum — Color Foundations
- Nielsen-Norman — Table of Contents Best Practices (2023)
- research-ops onboarding pattern:
research-ops/CLAUDE.md§8 - Brand palette math source:
marketing/landing/skills/landing/scripts/brand_palette_validator.py