| name | design-md-generator |
| description | Generates a DESIGN.md design-system document for the current project following the Stitch DESIGN.md format popularized by getdesign.md (VoltAgent/awesome-design-md). Use when the user asks to "create a design.md", "generate design system docs", "write DESIGN.md", "bootstrap UI style guide", "document the design system for AI agents", or when starting a frontend project and no design spec exists yet. Can start from scratch via guided Q&A or clone an existing community DESIGN.md as a starting point. |
design-md-generator
Generates a DESIGN.md in the project following the Stitch DESIGN.md format (9 sections), which the getdesign.md ecosystem has made the de-facto convention for giving coding agents (Claude, Copilot, Cursor) a machine-friendly design-system brief.
A good DESIGN.md means: when the user says "add a settings page", the AI produces a page that matches the rest of the app out of the box — no color/typography/component drift.
When to invoke
- User asks to "create a DESIGN.md", "generate design system docs", "bootstrap a style guide for this project".
- User has a new frontend project and wants a single source of truth for UI decisions.
- User has an existing UI and wants to reverse-document the design system.
- Polly invokes this during the greenfield flow when frontend is part of the stack.
Two modes
Mode A — From scratch (guided Q&A)
Walk the user through the 9 sections of the Stitch format, proposing options at each step based on their product/brand.
Mode B — Clone from community
Start from a published DESIGN.md in VoltAgent/awesome-design-md as a base, then customize.
Community clone is a single command the user runs — we don't run it:
npx getdesign@latest add <name>
The repo at https://github.com/VoltAgent/awesome-design-md has 68+ curated systems. After running the command, the user has a file to edit; skill-driver then guides the customization.
Output location
Default: docs/design.md (or docs/DESIGN.md if the project convention uses uppercase). Confirm with the user before writing — some projects prefer DESIGN.md at repo root.
The 9 sections — canonical Stitch format
For each section, produce concrete content, not placeholders. If the user can't answer a question, pick a sensible default aligned with their brand/tone and mark it "(default — adjust if needed)".
1. Visual Theme & Atmosphere
- Mood in 3-5 adjectives (e.g. warm, editorial, calm, rooted, deliberate).
- Density: tight / comfortable / generous (pick one).
- Design philosophy in 2-3 sentences. Reference a precedent if useful ("inspired by the NYT editorial layout", "Notion meets Dieter Rams").
2. Color Palette & Roles
Table with columns: Name · Hex · Role. Typical rows:
background / surface / surface-muted
text-primary / text-secondary / text-muted
accent (primary brand) / accent-hover / accent-muted
success / warning / error / info
border / border-subtle
Provide semantic names first, hex second. Light + dark variants if both are supported.
3. Typography Rules
- Font families: display + body + mono. Specify fallback stacks.
- Hierarchy table: Level · Font · Weight · Size · Line-height · Letter-spacing · Use for.
- Typical levels:
display, h1, h2, h3, body-lg, body, body-sm, caption, code.
4. Component Stylings
Minimum set: Button, Input, Card, Navigation, Modal/Dialog. For each, describe the default / hover / active / disabled states with color + border + shadow + padding.
5. Layout Principles
- Spacing scale (e.g.
4 / 8 / 12 / 16 / 24 / 32 / 48 / 64).
- Grid: max-width, columns, gutter.
- Whitespace philosophy (1-2 sentences).
6. Depth & Elevation
Shadow tokens with purpose:
shadow-sm (subtle lift — dropdowns, hover states)
shadow-md (modals, popovers)
shadow-lg (command palettes, overlays)
- Optional: border-based elevation for flat designs.
7. Do's and Don'ts
Two lists, 5-8 items each, phrased concretely:
- Do: "Use
accent only for the primary CTA on a page", "Keep body text at 16px minimum", "Prefer 1-2 weights per section".
- Don't: "No gradients on backgrounds (reserved for illustrations)", "Never put two
accent buttons side-by-side", "Avoid all-caps body copy".
8. Responsive Behavior
- Breakpoints (e.g.
sm: 640, md: 768, lg: 1024, xl: 1280).
- Touch targets (≥ 44×44 on mobile).
- Collapsing strategy: what hides on mobile, what switches to vertical stacks, what remains sticky.
9. Agent Prompt Guide
Short-form reference the AI agent reads before generating UI code:
- Quick color cheatsheet (just name → hex, no role)
- 3-5 canonical prompts: "a settings page in this design system", "an empty state illustration placeholder", "a destructive action confirmation".
- Common gotchas (e.g. "this design uses sans-serif for body; never switch to serif except for long-form reading views").
Workflow
- Confirm mode: scratch (A) or clone (B). If uncertain, suggest B if the user can name a brand whose style they like.
- Location: propose
docs/design.md, confirm.
- Phase 1 — Brand brief (skip in mode B if the clone is already aligned):
- Product name + 1-sentence description
- Audience (who uses it, tech comfort level)
- Tone (3-5 adjectives)
- Reference brands / sites the user admires (1-3)
- Phase 2 — Draft sections: produce sections 1-9 in order. For each, show a draft, ask the user "accept, refine, or skip?" before moving on.
- Phase 3 — Write: assemble into the final file, format as clean markdown (no HTML unless needed), save to the chosen path.
- Phase 4 — Update project meta:
- Add
design.md to nonoise.config.json under artifacts.design (if the config exists).
- Reference it from
CLAUDE.md / .github/copilot-instructions.md so agents read it when generating UI.
Template snippet (minimal example)
The following is a skeleton Claude/Copilot can fill with section 1 content:
# <Product Name> — Design System
> Source of truth for UI decisions. Agents (Claude Code, Copilot) should
> read this before generating any user-facing code.
## 1. Visual Theme & Atmosphere
**Mood**: warm, editorial, calm, rooted, deliberate.
**Density**: comfortable.
**Philosophy**: The product is a long-form reading tool for policy professionals.
The UI should feel like a thoughtful magazine, not a dashboard — generous
whitespace, one dominant typeface, accent color used sparingly for signals
(saved items, unread, destructive actions).
Repeat the same pattern for sections 2-9 with real content filled in from the Q&A.
Integration with other NONoise skills
frontend-design (bundled): produces distinctive, production-grade UI.
Once DESIGN.md exists, frontend-design must consult it before picking fonts,
colors, or layout patterns. The Agent Prompt Guide (section 9) is the handoff.
impeccable/* (vendored): the 18 impeccable skills (audit, polish, adapt,
etc.) can use DESIGN.md as the reference when critiquing existing UI.
- Polly: during the greenfield flow, propose running this skill right after
the user says "the frontend should look like X".
Rules
- Write concrete content, never placeholders. If the user can't decide, pick a sensible default and flag it
(default — revisit).
- Stitch format compliance: all 9 sections present, in order, with H2 heading per section.
- No proprietary brand imitation: the skill will not generate a system that explicitly copies a trademarked look (e.g. "make it exactly like Stripe"). Use the community DESIGN.md clone for legitimate reuse of published open systems.
- Update
CLAUDE.md pointer: append a line in CLAUDE.md (and .github/copilot-instructions.md if present) telling future agents: "When generating UI, read docs/design.md first." Use a managed block with HTML-comment markers so re-runs are idempotent:
<!-- >>> design-md (managed by design-md-generator) -->
## design.md
When generating user-facing UI code, first read `docs/design.md` for colors,
typography, spacing, and component rules. Section 9 (Agent Prompt Guide) has
the quick reference for common prompts.
<!-- <<< design-md -->
Out of scope (v1)
- Generating design-token JSON/CSS variables from the produced
DESIGN.md (future work — can be a companion skill).
- Auto-scraping a live website to reverse the design (future; in v1 this skill asks the user to provide the reference manually).
- Multi-brand / white-label support in a single
DESIGN.md (one file per brand).