Prompt · June 20, 2026
Claude Design Storyboard
A storyboard example for working with Claude Design
Vector Databases & HNSW — Storyboard (Video 22)
1080×1350 · 30fps · \~53s · concept explainer. Source: how vector search works, indexed with HNSW — Hierarchical Navigable Small World graphs(Malkov & Yashunin, 2016/2018). Companion to the memory/caching videos. Central metaphor: a vector database is a MAP OF MEANING — every document becomes a point in space where nearness = similarity, so "find related" becomes "find the nearest points." But with a billion points, checking every one is hopeless. HNSW lays a transit map over the space: local streets at the bottom, express lines up top — so you zoom from anywhere to the right neighborhood in a few dozen hops.Recurring proxy: a "meaning space" field of clustered ink dots (puppy/dog, car/truck); a copper query PIN; thin distance lines; a layered graph of 3 stacked planes (dense bottom "local" → sparse top "express"); a copper traversal path of glowing hops descending the layers; teal nearest-neighbor dots (top-k) with a ring; a recall/speed DIAL (ef · M). Color grammar: copper = the query and the active hop / the thing being searched (the active element), teal = the found nearest neighbors / the win. Close convention: "save this" — a concept explainer, save-it close, no next-video tease.
SCENE 01 · hook — 175f (VD1Hook) VO: "Search engines used to match keywords. Modern ones search by meaning — turning everything into a point in space and finding what's nearest. The catch: do that across a billion points, fast enough to feel instant." IMAGE: at f0 a copper query PIN is ALREADY dropped into a galaxy of thousands of ink dots, and the naive search is firing distance-lines to EVERY dot at once (a chaotic copper starburst), a counter spinning up "checking 1,000,000…". At \~0.4s it CHOKES — the starburst jams red, a clock pegs (thud + shudder ring). Title STAMPS on the choke: display "search by / meaning." Eyebrow + kicker trail after. PAYOFF: the query pin buried in a million distance-lines, the counter jammed red. Eyebrow "how vector databases work"; kicker "embeddings · ANN · HNSW."
SCENE 02 · re-hook · meaning becomes coordinates — 260f (VD2Embed) \[RE-HOOK\] VO: "It starts with embeddings. A model turns each piece of text or image into a long list of numbers — a point in a huge space — placed so that similar meanings sit close together. 'Puppy' lands near 'dog,' far from 'budget.' Now 'find related' just means 'find the nearest points to my query.'" IMAGE: words/images flow into a copper ModelChip and come out as VECTORS (short number strips) that FLY to positions in a 2-D meaning field, settling into labelled clusters (a "puppy/dog/ wolf" cluster forms tight; "car/truck" far away; "budget" off in another corner). A copper query PIN drops in and a soft radius highlights the dots nearest it (teal). PAYOFF: a clean clustered meaning-field, the query pin with its nearest dots glowing teal. Tag: "similar meaning = nearby point."
SCENE 03 · the brute-force wall — 250f (VD3Brute) VO: "But 'nearest' is expensive. The exact way is to measure the distance from your query to every single stored point and sort them. Fine for a few thousand. Hopeless for a billion — you'd be comparing forever. You need to find the neighbors without checking everyone." IMAGE: the field multiplies into a dense swarm; the query PIN fans a distance-line to EVERY dot, one after another, a "comparisons" counter racing toward "1,000,000,000" and a clock filling up copper, then alarming. Mono: "exact kNN = O(N) · compare to all." PAYOFF: the query drowning in distance-lines to a billion dots, the clock pinned. Tag: "checking everyone doesn't scale."
SCENE 04 · HNSW · build the layers — 250f (VD4Layers) VO: "The most popular fix is a graph called HNSW. Picture a transit map. The bottom layer connects every point to its nearby neighbors — local streets. Above it, sparser layers keep only a few points but add long express links across the whole space. Each level up has exponentially fewer stops." IMAGE: the flat field lifts into THREE stacked translucent planes (z-parallax). Bottom plane — every dot, wired to near neighbors (a dense copper proximity web, "local streets"). Middle — fewer dots, medium links. Top — only a handful of dots joined by long copper EXPRESS edges spanning the whole space. A few "elevator" links connect the same node across planes. Mono: "dense bottom · sparse top · \~log levels." PAYOFF: the three-layer graph — dense local web below, sparse express lines above. Tag: "a skip-list for space."
SCENE 05 · HNSW · the search hops — 250f (VD5Hops) VO: "A search starts at the top, on the express lines, and greedily hops toward whichever neighbor is closest to your query. When it can't get any closer, it drops down a layer and zooms in. Big jumps first, fine steps last — until it lands in the right neighborhood at the bottom." IMAGE: the copper query PIN enters at the TOP plane's entry node and HOPS along express edges, each hop snapping to the neighbor nearest the query (a glowing copper path, big strides). When a hop can't improve, it drops through an elevator link to the next plane (a satisfying ZOOM-in), hops again — strides shortening — then a tight local search on the bottom plane. \~12 glowing hops total, not a million. PAYOFF: the copper path threading top→bottom in a few big-then-small hops onto the target cluster. Tag: "greedy hop · descend · zoom in."
SCENE 06 · the payoff & the dial — 240f (VD6Win) VO: "Instead of a billion comparisons, it takes a few dozen hops — roughly logarithmic, often under a millisecond. It's approximate, not exact — but a single dial trades a hair of accuracy for speed, and recall is typically over ninety-nine percent." IMAGE: the landing neighborhood resolves — the true top-k dots light TEAL with a ring ("nearest neighbors"). The billion-counter from S3 collapses to "\~30 hops"; the clock snaps to "<1ms" (teal). A recall/speed DIAL (labelled "ef · M") sweeps: toward speed (faster, recall 98%) ↔ toward accuracy (recall 99.9%). Stamp: "billions → top-k in \~log steps · <1ms." PAYOFF: teal top-k neighbors found, the dial poised at ">99% recall," sub-millisecond clock. Tag: "approximate — but basically exact."
SCENE 07 · close — 175f (VD7Close) VO: "That's the engine under semantic search, recommendations, and RAG. Turn meaning into a map — then never read the whole map." IMAGE: bookend — the calm meaning-field with its layered graph faint beneath it; the copper query PIN traces one clean path down the layers to a teal cluster of top matches, then settles. Three small labels fade up: "search · recommend · RAG." Lockup: eyebrow "approximate, basically exact" + display "navigate / meaning." PAYOFF: kicker "save this." Wordmark.