design-brainlisted
Install: claude install-skill yapeepee/agent-harnesses
# Design Brain — 商品故事 ↔ 設計語言 grounded taste retrieval & composition
Evidence-grounded retrieval AND composition over **measured** design data — computed tokens, coherent recipes, hero geometry, motion, norm-deviations, screenshots, user verdicts — linked to brand-story attributes. v3 upgrades the system from a **point database** (sites as observed points, `/brief` as nearest-neighbor) toward a **factored space**: axes are grouped, gaps are computable, combinations have a grammar, and wholeness is verified round-trip.
v3.1 is v3's doctrine on repaired instruments. Every measurement fix below was written by the corpus about itself, in `ledger.md`. Read that file before changing this one: it is the only place the system records what its own eyes get wrong. The governing failure mode of an instrument is not noise — it is **reporting `0` when it means "I cannot see"**, because a fabricated zero is indistinguishable from evidence and will later be promoted to a correlation.
v3.2 extends v3.1 sideways, not upward, and assumes the corpus **grows without bound**: (i) the Structure group gets instruments — layout fingerprint + layout-dimension outliers — **fixture-gate PENDING** until `FIXTURES.md § v3.2` passes; (ii) the story side gets the same evidence lifecycle the design side always had (`story/registry.md` — folksonomy in, promotions earned, no ontology frozen at small N); (iii) `/brief` selects instead of blending; (iv) `/synthesize` gains tag-lifecycle upkeep, motif mining,