design-research

[SKILL] design-research [OBJ] Ground every design decision in real shipped-product evidence via the refero and mobbin MCP servers — then hand off to design-md-lord for token application.

[TOOL-MAP] Pick the layer that answers the question:

[RULES]

1. [PROHIBIT] No design from vibe memory — every major visual, layout, content, or interaction decision MUST trace to a searched reference, the user's brief, or a craft doc. Research before implementation, always.
2. [REQ] Search several angles before choosing: 3–5 queries — one broad aesthetic, one domain/category, one named strong product. Compare ≥3 real examples; never copy a single reference.
3. [REQ] Pick ONE dominant direction, don't average. When references conflict, keep the primary's signature traits sharp; secondary references contribute 1–2 narrow details only. Averaged references = generic AI slop.
4. [REQ] Evidence, not proof: a shipped example shows a product USED a design — it does not prove the design performs. A recurring pattern is a convention, not a best practice. Separate "observed on screen" from "recommended".
5. [REQ] Preserve provenance: for every load-bearing reference record product name + canonical URL (Refero/Mobbin link). Check capture metadata when recency matters.
6. [REQ] Token-meaning discipline: if a reference assigns a color/radius/shadow/role, use it only for that role — same law as design-md-lord rule 6–7.
7. [REQ] Brief before searching — missing info only if it materially changes the result: Designing [WHAT] for [WHO] on [PLATFORM]. Goal: [...]. Tone: [...]. Objection: [...]. Must remember: [HOOK]. Constraints: [...]. Research needed: [styles/screens/flows/sections].
8. [REQ] Reference-lock before build: lock (a) a user-provided source, (b) an existing design-system target, or (c) an explicit reference-locked direction + decision ledger (why this reference, what we take, what we reject).
9. [REQ] Validate after building: compare rendered output against the locked reference; fix design drift — research alone is not "done". Follow with DesignSlopVerifier + ui_a11y_checker (a11y is evaluated independently — a real product's choice is not an a11y guarantee).
10. [PROHIBIT] No verbatim copying: no protected layouts, branding, copy text, logos, or assets — extract the structure and the why, rebuild it for THIS product.
11. [REQ] Rate limits: Mobbin MCP = 60 req/60s — batch sensibly, respect Retry-After on 429. Weak results → query more specifically (platform + category + pattern + goal); too narrow → drop constraints.
12. [REQ] MCP absent? Do NOT fake research — state it plainly and proceed with design-md-lord's bundled catalog references + the user's own examples. Reference-lock still applies.

[WORKFLOWS]

1. New page/screen: brief → search_styles (direction) + search_screens/search_sections (composition) → compare 3+ → lock primary + ledger → design-md-lord for token application → build → slop + a11y verify.
2. Flow task (onboarding/checkout/paywall/cancel): search_flows → get_flow → map step goals/actions/system responses → adapt to product → build → verify sequence against reference flow.
3. Design critique/audit: capture current state → pull comparable search_screens evidence → report observed-vs-recommended per rule 4 → diff → fix.
4. Mobile: platform=ios (Refero) or Mobbin screens → respect platform HIG conventions found in references.