Skip to content

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

4 Commits
 
 
 
 
 
 
 
 

Repository files navigation

DesignOS

A complete design operating system for AI agents. Takes a product from raw idea to dev-ready design through five gated phases — with senior-level craft encoded, contrast mathematically verified, and a self-critique protocol that runs before you ever see the output.

Built for designers and founders who want designed products, not template output.

Install

npx skills add ravipalwe/designos

On the marketplace: skills.sh/ravipalwe/designos

Works with any agent that reads SKILL.md — Claude Code, Cursor, Windsurf, Amp, Devin, and friends. Then just say:

run DesignOS on my app idea

What it does

Brief → Direction → Design System → Screens → Handoff
Phase You get Gate
1 · Brief Taste profile (5 axes), users, top tasks, constraints quick confirm
2 · Direction 3 genuinely distinct directions rendered on your own product in one HTML file — pick one, mint 3 design principles + a signature element hard gate, never skipped
3 · Design System Full token spec (OKLCH-built color ramps, type matrix, spacing, motion) as YAML + a live HTML style guide, every contrast pair verified review & refine
4 · Screens One self-contained HTML mockup per screen — real content, real states, wired focus/hover, built only from system tokens per screen
5 · Handoff Token export (CSS light-dark() / Tailwind @theme / DTCG JSON), dependency-ordered build checklist, acceptance criteria, design-QA loop done

Everything lands in a design/ directory. design/STATE.md tracks every decision, so you can stop mid-process and resume days later — in a fresh session — with zero re-explaining.

Why it's different

Verified, not asserted. Ships with scripts/color.py (zero dependencies): WCAG 2.1 + APCA contrast checking with auto-fix suggestions, perceptually-even OKLCH ramp generation, hex→OKLCH inspection. The skill is forbidden from claiming contrast it hasn't computed.

$ python3 scripts/color.py check "8A8580 FAF7F2 muted"
label  fg       bg       WCAG    grade  need     APCA Lc  need    apca  verdict  suggestion
muted  #8A8580  #FAF7F2  3.42:1  AA-lg  4.5 req  +59.5    75 req  low   FAIL     try fg #67625D

(That failing pair was in this skill's own first draft — the script caught it before shipping. That's the loop working.)

Anti-generic by design. A taste-calibration step, 12 named style worlds with real recipes, an anti-slop scan targeting the recognizable AI-median look (Inter + purple gradient + glass cards), and a "one signature element" discipline so every product has the thing you describe to a friend afterward.

Crit before present. A 10-pass critique protocol — squint test, alignment walk, rhythm check, state completeness, slop scan, a11y pass — runs on the agent's own output before the human sees it.

Craft encoded, loaded just-in-time. Six reference files carry the knowledge (OKLCH hue-drift ramps, concentric radii, two-layer tinted shadows, type-scale matrices, motion choreography, content-first screen design) and are read only at the phase that needs them — the skill stays lean in context.

Mockups you can interrogate. Every screen ships with a dev overlay: press g for an 8px grid, b for a live squint test. Self-contained HTML — opens straight from disk, no build step.

Package

skills/designos/
  SKILL.md                  the conductor — phases, gates, modes, state
  references/
    direction.md            taste axes · 12 style worlds · anti-slop protocol
    foundations.md          color/OKLCH · typography · spacing · depth
    motion.md               duration & easing tokens · choreography
    screens.md              content-first method · real-data rules · craft bar
    critique.md             the 10-pass crit protocol
    handoff.md              token exports · build order · design-QA loop
  scripts/
    color.py                WCAG 2.1 + APCA + OKLCH — stdlib only, no deps

Modes: fresh start · resume · revision (names downstream cost first) · redesign of an existing product · brand intake (builds around your existing brand) · fast mode ("just run it").

Example

examples/solvent/ — DesignOS run on a demo product (Solvent, a runway tracker for bootstrapped founders). Open design/02-direction.html to see the direction gate: three worlds rendered on the product's own dashboard, contrast grades printed in each footer.

Requirements

  • File-system access (writes the design/ directory)
  • python3 for the color script (no packages needed)
  • Optional: Figma MCP for reading Figma references; browser tooling for visual self-verification

Security

  • scripts/color.py is pure stdlib — no network calls, no file writes, no dependencies to audit.
  • Like every agent skill, DesignOS runs with your agent's permissions. It's all readable markdown — read it before you run it.

License

MIT © Ravi Palwe

About

DesignOS — end-to-end design skill for AI agents: idea → direction → design system → screens → handoff, with verified contrast (WCAG+APCA) and anti-generic craft. Install: npx skills add ravipalwe/designos

Topics

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages