DESIGN.md is a structured markdown file that defines your project's brand spec — tokens (colors, typography, spacing, radius), voice, narrative, principles, personas, states, and motion. AI coding agents (Claude Code, Codex, OpenCode, Cursor) read it as authoritative context before generating any UI, so the output looks like your brand instead of generic 'modern minimalist' slop.

How does oh-my-design work?

Run `npx oh-my-design-cli install-skills` once in your project. It installs 17 skills (core flow + live-capture + asset + the v0.2 agent layer: orchestrator, kr-writer, locale-adapter, designer-review, final-qa, codex-image), 16 sub-agents (master orchestrator + 15 specialists), 4 hooks, and 326 reference DESIGN.md files. After restarting your AI coding agent, you just talk in natural language — skills route to the right sub-agents automatically.

Does it call any AI provider during install?

No. The install is a pure markdown copy — zero API calls, zero data leaves your machine. Your AI coding agent (Claude Code, Codex, OpenCode, or Cursor) is the inference layer; install just teaches it where to look.

Which AI coding agents are supported?

Claude Code, Codex, OpenCode, and Cursor — Cursor has a dedicated install channel (--agent cursor) that writes the .cursor/rules shim plus the shared reference catalog. The skill markdown files are agent-agnostic; hooks ship for Claude Code specifically.

What is the design harness?

The omd-harness skill runs a 10-phase design pipeline (Discovery → Research → IA → Wireframes → Design system → Components → Asset sourcing → Microcopy → Validation → Handoff). It dispatches sub-agents in parallel where independent, asks 3 mandatory user checkpoints, and emits a v0/Cursor-ready package. Trigger it by saying things like 'design the entire onboarding from scratch' or '/omd-harness '.

Yes — MIT licensed, open source, no signup, no API key, no paid tier. The references in the bundle belong to their respective companies and are reproduced for educational reference only.

oh-my-design v1.8.5 (2026-06-23)

QA pass on the 2026-06-22 batch: 40 refs re-inspected vs live + benchmarked against the designlang deterministic extractor. 35/40 PASS, mean accuracy 98.3/100, 5 MINOR token-value fixes applied.

Two multi-agent workflows: per-ref live re-inspection + adversarial verify (QA), and our-tokens-vs-designlang-vs-reality comparison. Zero major errors, zero fabrications; brand primaries/fonts all held. designlang independently corroborated ~5.4/6 of our key tokens per ref.
5 confirmed MINOR corrections (structural token values, not brand errors): paypal card shadow 0.1/4px/12px → 0.08/24px/48px (live layered-card); chunghwa radius scale +20px (dominant modal/card default, ~46%); hyundai chatbot FAB box-shadow rgba(0,0,0,0.15) 0 0 20px added (CREATE sampled the wrapper, not the button); snapchat §6 "none across nav" softened — components flat but marketing nav wrapper carries one shadow; octopusenergy postcode input split into hero (white/Arial-600) vs tariffs (transparent/Chromatophore) surfaces. All 5 verify-reference green; 553 tests pass.
Benchmark write-up: data/reference-audits/2026-06-23-qa-designlang-benchmark.md. Finding: our LLM pipeline is 39–0 (1 tie) more brand-accurate than designlang (which picks most-frequent color as "primary", ~35% brand-correct); recommendation = fold designlang in as a free deterministic Stage 0 (palette histograms + DTCG) feeding our brand-judgment pass.

QA pass on the 2026-06-22 batch: 40 refs re-inspected vs live + benchmarked against the `designlang` deterministic extractor. 35/40 PASS, mean accuracy 98.3/100, 5 MINOR token-value fixes applied.