Perfect ai design

AI Design OS v3

Purpose: stop AI-generated UI from looking AI-generated, and give it a real direction. Bans raise the floor; a committed direction raises the ceiling; the auditor enforces both.

Core loop (works even with zero tool calls)

1. Design Read — one line before any code: Reading this as: <page_kind> for <audience>, with a <vibe> language, leaning toward <direction/system>. Generic "modern SaaS landing" = fail. Ask ONE question if the brief is truly ambiguous.
2. Dials — declare DESIGN_VARIANCE / MOTION_INTENSITY / VISUAL_DENSITY (1-10). Never silent defaults.
3. Commit ONE direction + locks: accent (one), type display+body, radius system, spacing scale, theme, ONE signature element. Hybrid directions are invalid.
4. Generate under locks. Surface classes differ: marketing gets hero laws; dashboards get density/scanability laws (no poster heroes); slides get slide floors.
5. Audit — run the audit tool on generated files. Fix all hard findings; fix default findings or waive each with a one-line justification.
6. Subject-fit test — would these tokens paste onto another product unchanged? Must be NO. Then hand off: frontend-design (visual polish), designdna (tokens/system).

Tools (MCP ai-design server, or CLI fallback)

Tool	When	CLI
plan	any multi-step design task, call FIRST	ai-design plan "<task>"
contract	before generating UI	ai-design contract "<brief>"
direction	commit a positive direction (14 cards, tokens included)	ai-design direction pick --brief "<brief>"
audit	AFTER generating: scan files for AI tells (file:line evidence)	ai-design audit <paths>
visual	score a screenshot: AI-face / craft / authorship (host fills rubric; or set AI_DESIGN_VISION)	ai-design visual <img> --code ...
checklist	structural checks greps can't catch; visual = screenshot scoring protocol	ai-design checklist de-ai
search	stack picks, real citations	ai-design search "<q>"

No MCP and no CLI? Read directly: catalog/taste_contract.json (bans + limits), catalog/directions.json (direction cards), references/checklists/de-ai.md (gate A-K), references/checklists/visual-critique.md (screenshot scoring: AI-face / craft / authorship).

Hard rules

1. Contract before pixels; audit after pixels. Never skip both.
2. hard audit findings are non-negotiable. default findings need the user's brief to justify them, not your preference.
3. One direction, one accent, one signature. The signature comes from the direction card applied to THIS subject.
4. Dashboards/utility surfaces: official design-system packages beat hand-rolled CSS; no landing-page theatrics.
5. Redesigns: audit existing tokens/IA first; never silently change slugs, nav labels, form fields, logo, legal copy, analytics IDs.
6. Cite catalog ids/urls from search; never invent papers, stars, or links.
7. User can't articulate the style? NEVER ask open-ended "what style do you want". Use direction(action='ask'): 2-3 binary forks with concrete picture descriptions. Abstract adjectives (高级/简约/大气/科技感) → direction(action='vocab', term=...) for anchored readings + the anti-interpretation. Still stuck after 2-3 forks → generate hero-only contrast stubs from the top-2 direction cards and let the user point.
8. Novelty is not the goal. Taste is refusing the mean.