A markdown template system that turns any AI agent into a reliable premium-grade product builder.
Brand identity slots into ~30 fields per project. The rest — design tokens, component specs, page patterns, voice rules, accessibility floors — is pre-decided based on what premium design teams actually do. Hand the filled-in files to any AI tool (Claude, Cursor, ChatGPT, Cody, others) as the single source of truth.
The system is built from ~4,000 lines of cross-referenced research on what makes premium product design measurably different from generic SaaS-template output. It's the thing I wish existed before I started building it.
- Markdown templates covering brand context, design system, content / layout, and orchestration
- Web and mobile treated as parallel systems with shared brand identity
- Brand-agnostic — works for any project; you fill in slots
- Free, MIT licensed, provided as-is — fork, adapt, ship anything you want
- Compatible with any AI tool that reads markdown
The ten templates live in templates/. (The release download flattens them into one folder — instance files are flat siblings.)
| File | Purpose | Lines |
|---|---|---|
PROJECT_TEMPLATE.md |
Entry-point orchestration — tells the AI which sibling files exist and in what priority. Lightweight index. | ~300 |
INFORMATION_TEMPLATE.md |
Brand identity, audience persona (with anti-personas), business model, voice principles, product features + non-features, social, SEO defaults. The "why" and "who." Shared between web and mobile. | ~500 |
CONTENT_TEMPLATE.md |
Reusable content model — typed records (testimonials, stats, features, FAQs, team, case studies, integrations, pricing) the AI maintains once and pages reference via {content.*}. Single source of content truth; no drift across pages. Shared web + mobile. |
~130 |
DESIGN_TEMPLATE_WEB.md |
Visual design system for web — colors (OKLCH), typography, spacing, ~25 component specs, motion, accessibility (APCA + WCAG 2.2). | ~2,300 |
DESIGN_TEMPLATE_MOBILE.md |
Visual design system for mobile — iOS HIG + Material 3 native specs, gestures, haptics, safe areas, Dynamic Type / sp scaling. | ~1,800 |
SPEC_TEMPLATE_WEB.md |
Site map + per-page content/layout/copy + forms + system messages + transactional email + legal pages + analytics events. | ~950 |
SPEC_TEMPLATE_MOBILE.md |
App map + per-screen content/copy/states + onboarding + auth + permission pre-prompts + push notifications + app store metadata. | ~860 |
SEO_TEMPLATE.md |
Discoverability — SEO + AEO + GEO: JSON-LD structured data (page-type schema map), answer-first content patterns, generative-engine citability (stats / quotes / citations / entity consistency), freshness, social cards, llms.txt. The discoverability layer over INFORMATION.seo + SPEC. |
~130 |
QA_TEMPLATE.md |
Premium acceptance gate the AI runs against its own build before "done" — WCAG 2.2 AA (axe zero-critical), Core Web Vitals budget, token fidelity (no hardcoded values), responsive incl. ultra-wide, content-matches-SPEC, security. | ~150 |
DECISIONS_TEMPLATE.md |
Decision log — an append-only ADR-style record of locked choices + rationale. A fresh agent reads it before proposing changes, so settled decisions aren't relitigated each session. | ~70 |
Plus one reference file:
| File | Purpose |
|---|---|
research.md |
The brand-agnostic premium-standard research (29 sections, ~4,090 lines) covering design, copy, discoverability, and the build process (incl. §X2 Marketing & Conversion Copy, §X3 SEO + AEO + GEO, and §X4 Build Modes / reference-driven design / visual iteration) — the explanatory backing for everything in the templates. DESIGN / SPEC / SEO templates cite specific sections. |
Plus optional tooling (the markdown templates stay dependency-free):
| Folder | Purpose |
|---|---|
tools/brand-kit/ |
An OKLCH palette generator (model-only) + a brand-kit viewer that reads a project's DESIGN.md + INFORMATION.md and renders the whole design system (specimen + composition, light/dark, the project's real fonts + icon family). Run npm run dev to review before building. See tools/brand-kit/README.md. Note: the tooling lives in the repo — clone it to use the generator/viewer. It is intentionally excluded from the lean release download (the markdown templates are self-contained and can be handed to any AI without it). |
Most AI tools, given a prompt like "build a pricing page," regress toward generic SaaS output — because they average across millions of mediocre examples in their training data. Without an anchor, even capable models produce drift between sessions, between tools, and between pages of the same project.
These templates are that anchor. They encode:
- Universal rules that don't change between projects (accessibility floor, anti-patterns, scale systems, the AI Agent Contract)
- Structural defaults representing the broad-premium-middle of what mature design teams choose
- Brand-specific slots you fill in per project
The AI now references concrete tokens ({colors.primary.9}, {spacing.scale.4}, {typography.roles.body-md}) — and reusable content the same way ({content.testimonials.maria.quote}, {content.stats.active_teams.value}) — instead of guessing. Every session lands on the same answer. Cross-page, cross-tool, cross-team consistency becomes the default rather than a constant battle.
Each template is operational, not aspirational. Specifically:
- Colors — OKLCH authoring, 12-step Radix scale, APCA contrast targets, dark-mode strategy (perceptually mapped, not inverted), surface hierarchy
- Typography — 8 semantic type roles, modular scale ratios, line-height + tracking inverse rules (font-category-aware: serif / sans / mono), variable fonts, fluid
clamp()sizing - Spacing — 4 px base, 18-step scale, inset/stack/inline patterns, container queries
- Shapes — 6-step radius scale, nested-radius math, continuous corners on iOS
- Elevation & surface separation — 6-level layered shadows (web) / materials + M3 (mobile), plus a tunable separation system (shadow / border / surface-tone, shadow strength + size, border width)
- Motion — 5 duration tokens, 4 easings, 3 spring presets, FLIP, View Transitions API, reduced-motion variants
- States — 10 canonical states, focus-visible spec, touch states (iOS dim / Android ripple), disabled-without-opacity
- Iconography — 6 sizes, stroke-to-text-weight pairing, optical alignment, recommended free + premium libraries (Lucide / Phosphor / HugeIcons)
- Imagery — 6 aspect ratios, modern image loading (AVIF/WebP/JPG, srcset, fetchpriority)
- Accessibility — WCAG 2.2 + APCA contrast targets, all 6 preference media queries (including
forced-colors), live regions, skip links - ~25 components (web) / ~22 components (mobile) — pixel-perfect specs with state matrices for every common atom and molecule
- Web patterns — surface-scoped (marketing landing vs product web-app), canonical landing composition, hero variants, scroll-triggered animation framework, bento grids, command palette, code surfaces
- Mobile patterns — iOS HIG + Material 3 native conventions, gestures, haptics, safe areas, permissions pre-prompting
- Data visualization — three-palette system (categorical / sequential / diverging), chart-type conventions
- Internationalization & RTL — CSS logical properties, mirror rules, tall-script line-height, CJK exceptions, locale formatting
- Microcopy — voice principles, banned-word list, length budgets, premium positioning structure
- Marketing & conversion copy — positioning→messaging→copy, awareness stages, the conversion-page narrative, headline/value-prop formulas, feature→benefit→outcome, persuasion psychology (ethical), proof & CTA craft, app-store/onboarding/paywall copy, answer-engine (AEO/GEO) copy (
research.md §X2; SPEC templates apply it) - AI Agent Contract — 26 hard rules (web) / 18 hard rules (mobile) the AI must follow
- Content model (
CONTENT.md) — reusable typed records (testimonials, stats, features, FAQs, team, case studies, pricing) maintained once and referenced by pages via{content.*}— the same mechanism as design tokens, so the same quote or metric never drifts across pages - Discoverability (
SEO.md, backed byresearch.md §X3) — three layers: SEO (rank — crawlability, structured data, E-E-A-T, Core Web Vitals), AEO (be the direct answer in AI Overviews — answer-first, FAQ markup), GEO (be cited by ChatGPT / Claude / Perplexity / Gemini — stats / quotes / citations / entity consistency / freshness): JSON-LD per page type, social cards,llms.txt - QA acceptance gate (
QA.md) — the AI self-audits its build before "done": WCAG 2.2 AA (axe zero-critical), Core Web Vitals budget (LCP/INP/CLS), token fidelity, responsive incl. ultra-wide / 4K, content-matches-SPEC, security - Decision log (
DECISIONS.md) — append-only ADR-style record of locked choices + rationale, so a fresh agent reads what's settled before reopening it; plus anAGENTS.mdemitted to the built repo (the open standard 20+ coding agents read) for the docs→code handoff - Build process (
research.md §X4) — recognizes the real ways you build: docs-first → build, explore-build → design-off-it → document → rebuild, or hybrid; how to use uploaded reference images / sketches (translate their structure into your tokens, don't copy a brand); and the visual-iteration loop — drive a browser (e.g. Claude in Chrome) to view the running dev server and iterate until it looks right, not just compiles - Value-state tracking (
PROJECT.md §Status protocol) — every value carries a knowable state via a tiny inline-marker vocabulary:# draft(AI-inferred, needs your approval → intake PART 0),# default(out-of-box default you never chose → PART 2 review),# approved/# locked, and unmarked = your own input. A hard producer rule (the AI never writes an inferred value or accepts a default without its marker) plus a state ledger at every audit guarantee that nothing fabricated ever passes silently as approved truth
# 1. Get the core templates (they live in templates/). Or download the release zip for all 10, flat.
curl -O https://raw.githubusercontent.com/arshawnarbabi/premium.md/main/templates/PROJECT_TEMPLATE.md
curl -O https://raw.githubusercontent.com/arshawnarbabi/premium.md/main/templates/INFORMATION_TEMPLATE.md
curl -O https://raw.githubusercontent.com/arshawnarbabi/premium.md/main/templates/DESIGN_TEMPLATE_WEB.md
curl -O https://raw.githubusercontent.com/arshawnarbabi/premium.md/main/templates/SPEC_TEMPLATE_WEB.md
# Optional add-ons (same templates/ path, or let the intake create them):
# CONTENT_TEMPLATE.md · SEO_TEMPLATE.md · QA_TEMPLATE.md · DECISIONS_TEMPLATE.md
# 2. Rename to project files. PROJECT.md is the generic entry-point name; you can
# rename it to CLAUDE.md / AGENTS.md / etc. to match your AI tool (see "Tool conventions" below).
mv PROJECT_TEMPLATE.md PROJECT.md
mv INFORMATION_TEMPLATE.md INFORMATION.md
mv DESIGN_TEMPLATE_WEB.md DESIGN.md
mv SPEC_TEMPLATE_WEB.md SPEC.md
# 3. Fill the brand-identity slots. Greppable until clean:
grep -n "<[^>]*>" PROJECT.md INFORMATION.md DESIGN.md SPEC.md
# 4. For DESIGN.md, generate the color palette with the bundled generator:
# cd tools/brand-kit && npm run gen -- --base "#2D6A4F" (your brand HEX or oklch(L C H))
# (the generator/viewer live in tools/ — clone the repo to use them; or just ask your AI to
# derive the palette per §Colors → Generating the scale. No install needed for the generator.)
# 5. Hand all four files to your AI tool. PROJECT.md is the entry point.
# 6. Review the result in the brand-kit viewer: cd tools/brand-kit && npm run devSame pattern but with DESIGN_TEMPLATE_MOBILE.md → DESIGN_MOBILE.md and SPEC_TEMPLATE_MOBILE.md → SPEC_MOBILE.md.
Use all the relevant files — PROJECT, INFORMATION, CONTENT, DESIGN ×2, SPEC ×2, SEO, QA, DECISIONS. The brand-identity slots in INFORMATION.md are shared — fill once. The visual / content templates split by platform.
Beyond copying the raw templates, the whole system ships as a cross-tool Agent Skill (SKILL.md, the open standard) — so your coding agent loads it automatically when you ask it to build or design a premium product, and reaches the templates + research + brand-kit on demand. The skill is self-contained (the reference material travels with it).
One command — installs into every detected agent (Claude Code, Codex, Cursor, OpenCode, …) via the skills CLI:
npx skills add arshawnarbabi/premium.md --skill premium-md
Then just ask — e.g. "set up the design system for my new app" — and the agent loads the skill; or invoke it directly with /premium-md.
By default this installs into the current project (./.agents/skills/ or ./.claude/skills/). Add -g to install it globally (available in every project):
npx skills add arshawnarbabi/premium.md --skill premium-md -g
Fallback (no Node, any tool) — clone and copy the skill folder into your agent's skills directory:
git clone https://github.com/arshawnarbabi/premium.md
cp -r premium.md/skills/premium-md ~/.claude/skills/ # Claude Code
cp -r premium.md/skills/premium-md ~/.agents/skills/ # Codex / Cursor
Either way you get the same thing the raw templates give you — just auto-loaded and kept together. The plain-template path above still works for any tool or for browsing.
You don't have to fill the templates manually. Once they're copied into your project, ask your AI:
"Help me populate these templates"
It produces a structured intake form — must-fill brand identity at the top, customizable defaults below — covering every decision you need to make. Answer in your own time. The AI fills the templates for you, derives the 12-step color palette + dark mode counterpart, propagates shared values across every template, and runs final verification. For a web + mobile project it confirms scope up front and fills both design files — shared values (colors, fonts, icons) are copied across automatically, so the mobile system never gets left behind.
Trigger phrases the AI listens for: "help me populate this" / "what do you need to know?" / "run the intake" / "walk me through this". The intake also returns a state ledger — empty slots in PART 1, AI-drafted values awaiting your sign-off in PART 0, and unconfirmed defaults flagged for review — so "what's left to fill out or approve?" has a complete, honest answer (see PROJECT.md §Status protocol).
The full Interactive Population Protocol lives in PROJECT.md. See it for the exact intake structure, behavioral steps, and cross-template consistency rules.
In your AI prompt, reference the entry-point file. The AI discovers the rest via PROJECT.md's declarations.
Use the templates in this project as your source of truth.
Start by reading PROJECT.md, then consult the files it declares in the
priority order it specifies. Reference tokens via {group.path} syntax;
never invent values not in the documents.
Now build me a [pricing page | onboarding flow | settings screen | …].
Tool conventions: rename PROJECT.md to match what your tool auto-discovers:
| Tool | File name |
|---|---|
| Claude Code | CLAUDE.md |
| Cursor | .cursorrules |
| Generic AI-agent tools | AGENTS.md |
| Continue.dev | .continuerules |
Content stays the same. AGENTS.md is an open standard (stewarded under the Linux Foundation) that 20+ coding agents read — Codex, Cursor, Copilot, Gemini CLI, Zed, and others. When the design docs are turned into a real codebase, the build step should also emit an AGENTS.md at the repo root carrying the build / test / lint commands, code conventions, the QA gate, and a pointer back to these design docs — so the coding agent that maintains the repo has its own entry point.
Required brand inputs (~10 fields — no defaults work): Brand name, description, audience, voice, brand primary color (OKLCH), brand neutral hue, display + body + mono font families, project type (marketing site / product SaaS / mobile app / hybrid).
Profile selections (9 profiles — all have premium defaults): Radius, type scale ratio, density, motion personality, elevation depth, color saturation, brand warmth, section padding, chart minimalism.
Pick-one slots (~22 web / ~14 mobile — all have defaults): Input style, tabs style, icon fill, avatar shape, modal backdrop, code surface, onboarding pattern, save model, settings IA, command palette, RTL support, chart library, illustration style, and others.
SPEC content (per-project, no defaults): Site map / app map, page sections, copy, voice samples, forms, notifications, email templates, app store metadata.
INFORMATION content (per-project, no defaults): Audience persona depth, market positioning, business model, brand story, social handles, legal jurisdiction.
Total decisions to fully configure a project: ~30 quick decisions + writing the actual content/copy. Most decisions take seconds; the content takes real time (as it should).
| Tier | What | How it appears |
|---|---|---|
| Universal | Hardcoded; immutable per project | "Spacing values must be multiples of base"; "Use APCA for contrast"; "Touch targets ≥ 44 pt mobile" |
| Structural default | Pre-filled, sensible, overridable | "Type scale ratio: balanced (1.200)"; "Modal backdrop: blur"; "Onboarding: empty-state-driven" |
| Brand-specific | Empty <slot> per project |
Brand colors, font families, audience, voice, page content |
| Concern | Default |
|---|---|
| Framework (web) | React + Next.js (App Router) |
| Styling | Tailwind v4 |
| UI primitives | shadcn/ui (dashboards / product apps); custom for marketing |
| Animation | Framer Motion |
| Forms | React Hook Form + Zod |
| Icons | Lucide (free default) / Phosphor / Heroicons / Tabler / HugeIcons (free Stroke-Rounded ~4,500 + 51K on Pro) — declare per project; keep PROJECT.tech.web.icons in sync with DESIGN.icons.library |
| Deploy | Vercel |
| Mobile native | SwiftUI / Jetpack Compose, or React Native (Expo) for cross-platform |
research.md documents the source for every rule. Key references:
- Published premium design systems: Radix Colors, Material 3, Apple HIG, IBM Carbon, Atlassian Design System, Shopify Polaris, Geist (Vercel)
- Standards: DTCG W3C Design Tokens Format Module (stable Oct 2025), OKLCH color, APCA contrast, CSS logical properties, View Transitions API, container queries
- Practitioner content: Karri Saarinen's 10 rules of craft, Rauno Freiberg's interaction principles ("Devouring Details"), Vercel's published "Web Interface Guidelines"
- Production-site analysis: Linear, Vercel, Stripe, Notion, Anthropic, Mercury, and Pixel Point's case-study portfolio
The research file is informational — you don't need it to use the templates. It's the explanatory backing if anyone asks "why this rule?"
All templates carry template_version: "1.21.0" in their YAML frontmatter. Per-project instances should preserve this field — when the template family evolves, projects can track which version they were authored against.
This release: v1.21.0 — stable. Future updates follow semantic versioning.
The DESIGN.md format used here — YAML token frontmatter + a human-readable rationale body, with {group.path} token references and sections like Overview / Colors / Typography / Elevation / Shapes / Components / Do's & Don'ts — builds on Google Labs' DESIGN.md (Apache-2.0). This project extends that single-file format into a multi-file product system (brand, content, web + mobile design, discoverability, and a QA gate), adds an interactive intake protocol and OKLCH brand-kit tooling, and backs the rules with ~4,000 lines of research.
Issues and discussions welcome. PRs welcome but not promised. This is maintained as time permits.
If you ship something with this and want to share, drop a link in a discussion — I'd love to see what people build.
MIT © 2026 Arshawn Arbabi
Provided as-is. Free to use, fork, adapt, and ship anything you want. No attribution required (though appreciated). No warranty. No support obligation on my end — though I'll engage with the community when I can.
Built standing on the shoulders of public design system documentation from Radix, Tailwind, Material, Apple, Atlassian, Shopify, and Vercel. Practitioner principles from Karri Saarinen (Linear) and Rauno Freiberg (Vercel). The DTCG W3C Design Tokens working group. Every team that publishes their design system openly — you make work like this possible.