Finding references and assets

The roles a reference may play, where to search, and how to vet. scripts/find_refs.py does the fetching; run it with no arguments for the command list. Output is one JSON object per line.

Roles

Every reference gets exactly one role, a take (what it supplies) and a fidelity word. A file good for two roles is listed twice with two scoped takes.

Assets are not references. Models, fonts, footage, logos, placeholder photos and audio go under assets: with use as-is or one-to-one.

Handles use the role letter: m motion, a art direction, u ui, c copy, d detail, f family, t technique.

Order of work

References come first, before any beat (SKILL.md steps 2-4): five images from one aesthetic and three moving references, then, for all eight, the beat each feeds and what is taken. Five tiles from one channel is normal; two channels of the same aesthetic is fine; two aesthetics is not. For motion the test is "would he add this to mo-graph?", not "does it show the right object?": a dull move on the right object loses to a striking move on a different subject. Before falling back on a camera move written in words, try channels "camera move", "zoom", "transitions".

The fence: a named style or an attached reference

Blake names niche aesthetics himself; other users name mainstream ones in plain words (clean and minimal, lots of white, like stripe's style, like claymation). Both past failures (overriding the word, screenshotting it; see patterns.md) treated the fence as a place to copy from, not as a query. So:

1. Decode the word in one line: what the user pictures when they say it. minimal, lots of white = big type, one accent, generous margins; claymation = stop-motion clay, fingerprints, boil, on twos; like stripe, dark mode = navy, blurple, fine gradient lines, mono code. That line is the fence; it constrains palette, material, type and component vocabulary. It does not constrain layout, scale, camera or the mechanic.
2. The word is the are.na query. Still are.na, still curated, but the style or company name is the constraint. Two kinds of query, both needed:
   • The name itself, with search "<style word>" / search "<brand domain>" (the combined endpoint; channels only matches titles and stripe titles are fashion stripes). For a style word it returns the channels named for it plus the canon (claymation → six channels, Aardman, Bickford, Will Vinton). For a brand domain it returns the brand's own artefacts (stripe.com → the gradient wave plates, the letterbox, the dashboard redesign, the gradient codepen): the a1 look tile and the facts, inside the fence by construction. connections from brand blocks leads to dump channels; skip it for brands, use it for style words.
   • The decoded visual words, with channels "<word>": the fence turned into the two or three things a curator would collect. Stripe = gradient (gradient-airbrush 273/316, odd-gradient), glow, iridescent; minimal white = big type (big-type-container 1477, big-juicy-type), white space, swiss; claymation = clay, stop motion. These channels hold what is striking about the fence and supply a2-a5 (palette, material, composition, detail). A product-UI channel (fintech, minimal-ui, dashboard) is a u source only, never a1-a5: its tiles are screenshots and the build will be one too.
   • Umbrella channels stay a source inside the fence: umbrella <slug> "<word>|<decoded word>" finds his nested channels that carry it. Sheet two or three per kind, drop the homonyms by eye (stripe fashion, clay-play = clay.com), pick the tile that stops you. If the name query is all homonyms, the decoded-word and umbrella channels carry the fence on their own; a channel named for the subject (menu design for a menu clip) is not a way round it. Remarkable is judged among tiles that pass the fence; a tile outside it loses however striking.
3. The primary site supplies facts, not frames. For like <brand> take the hex, the typeface and one real component from the product surface (homepage, docs, app) and write them into a1/a2's take. A docs page, a settings screen or a pricing table is never a composition reference; the composition ref is a designed frame (a hero, a poster, a product-film still, a campaign still in the product palette). Campaign and event work is fine for composition and material when its palette is the product's; it was the Sessions palette, not its provenance, that missed.
4. Motion is never fenced. m1-m3 still come from mo-graph and its nested channels and still pass "would he put this in mo-graph?". UI micro-interaction channels are a ui source, not an m source. The look is applied to the move, not the move to the look; a flat look with a flat move is a screenshot.
5. An attached reference fixes the look. It takes the role its phrasing implies (an attached still is a1 unless the draft says otherwise), and the other four images come from the same look: connections <block_id> if it is an are.na block, else the channel or site it came from, else channels "<its aesthetic>". Never switch aesthetic because another sheet is stronger.
6. No fence (the draft says nothing about look, brand or medium): the open rule applies, pick for remarkable.

Then, as the beats need them:

• copy / assets: the subject is a specific recognisable thing (the user's own app, a named painting, product, train car, mascot, logo). Search the source and the assets together; for 3D, Sketchfab whenever the object is hard to model convincingly in code, even if the draft never mentions a model. If the thing is the user's own and nothing is attached, say so above the prompt and name the stand-in.
• ui: a beat shows an interface and there is no screenshot or earlier mock.
• detail: one beat needs a feature none of the five images show (a motif, a pattern for the turn beat).

If a mandatory slot is still empty after are.na, Pinterest and primary sites, write [? motion ref] / [? visual ref] into the prompt beside a words-only description and name the queries tried in the fill-ins. Never pass an unvetted hit to fill a slot.

Skip in delta, image-feedback and housekeeping modes unless the change brings in a new look or move. Skip roles the user already attached.

Search order

Images and motion: are.na channels → Pinterest → primary sites. 3D: Sketchfab first. Go to Pinterest only for a role are.na left without a passing candidate; when both pass, keep the are.na one. Primary sites when both are empty, and always for the final file of a copy reference.

are.na: channels, never block keywords

Block keyword search matches the text of the channels a block sits in, not what it shows (peel gif returned a blob, a leaf, a banana and a lily). The one exception is a fence word: search "<style>" / search "<brand domain>" (the combined endpoint) is how a named style or company is found, see "The fence". The loop, per role:

1. Seed from the author's channels (below). Most are umbrellas: open them with umbrella and work in the nested channel that fits, never the umbrella's own mixed blocks.
2. Widen. channels "<query>" lists channels whose title matches, ranked by followers, 20+ blocks. connections <block_id> [n] lists the other channels holding a good block (12 biggest by default): curators with the same eye, usually the best next stop. A query with no hits prints one "empty": true line.
3. Rank. Prefer titles naming the exact mechanic / aesthetic / ui type, 50-2,000 blocks, followers as a quality signal. Skip catch-all dumps however large.
4. Sheet. channel <slug> .context/refs/<topic>/arena 48 <moving|still> [regex] writes numbered contact sheets (24 tiles each). Read them. Under a third of the tiles fitting: drop the channel. The regex filters title / description / source, but most blocks are untitled, so the sheet is the real filter.
5. Pick tiles by number and vet (below). Note the channel next to the block in refs: so the user can browse the neighbours.

Stop after three channels per role, then go to Pinterest.

Motion seeds, always sheeted with moving:

• mo-graph-1uxjlkjqqhc (76 blocks, gifs and mp4s; pass the bare slug, blake-shao/... 404s)
• its five nested channels: TYPE ANIMATION REFERENCES (type-animation-references, 474), DESIGN // MOVING (design-moving-l9-xt3sxtne, 406), Motion Inspo (motion-inspo-tfrw-gidqfw, 205), AD | Device / Tech (ad-device-tech, 182), Product video (product-video-mf4kzs1zbo0, 22). Type beats → TYPE ANIMATION; device or product beats → AD | Device / Tech and Product video.
• widen with channels "<mechanic> animation", "motion design", "kinetic type", "transitions", and connections from the best mo-graph block (leads to 2D Motion GIF, design/motion, Movimento).

Art direction: the visual reference is the aesthetic. Do not pick a look in words and then search for a picture of it; browse the menu of aesthetics, sheet the two or three that could carry the subject, and let the tile that stops you decide. With a fence (above), the two or three sheeted are instances of the decoded line. If the brief names none and the file gives none, the menu is:

• Menu: the nested-channel titles of his umbrellas (umbrella <slug>): vis-4hvdpzvvua4 (73 nested, graphic and visual culture: Neo-gothic, Digital Embroidery, ASCII art, y2k playstation 3d stuff, Anime Computing, Baroque User Interface, Printed Matter, practical sigil magic...), flat-ovcn8k0nwa0 (13, poster / type / 2D: @swissposters, Japanese Graphic Design, Chinese Typography, Pattern, single-designer channels), form-glcjdzmigiq (15, objects: Pocket Tech, Handheld, Industrial Design, form follows fiction), mercenary (3, Visual Systems, Design Systems). Smaller umbrellas: yiwqez8yg-e (Japanese houses), translucent-skin, root-systems-m5flfkrwpe8, foundry-p8qvueq4coo. embroidery-jt4hrowbvw4 (19 blocks; sheet it, its one nested channel is Woven Images), data-geo, smart-shamate, b48e9jcrcse, baseball-hat-alphabet are flat single collections (families).
• Plus his past named styles (riso, frutiger aero, 16-bit, ukiyo-e, porcelain, embroidery, klein blue, PS2 low-poly) and CARI's ~240 named-aesthetic channels (user-channels evan-collins-1522646491: Y2K, Frutiger Aero, Vectorheart, Cassette Futurism, Whimsigothic...). Prefer one from his umbrellas.
• Fit it to the subject and the file; never blend two. On a bare draft avoid the safe default (flat vector on white, dark "linear style", brushed metal) unless a sheet makes it striking; his own openers never pick the safe default. When the user asked for that default by name, it is the fence and the job is the best version of it, not an escape from it. Put the chosen aesthetic in the prompt's first line (in neo-gothic) and give the one-line reason plus the runner-up aesthetic in the fill-ins.

Then sheet the channel that is that aesthetic, still: the nested channel from his umbrella, then the CARI channel of the same name, then channels "<aesthetic>". Pick five tiles with five jobs (look, palette, material, composition, detail). If the channel cannot supply five, sheet a second channel of the same aesthetic before switching aesthetic.

UI seeds, ui-specific channels only, brand-neutral take:

• his own: umbrella mercenary, and umbrella vis-4hvdpzvvua4 "interface|computing|apple|fidelity"
• layout: channels "ui design", "interface", "dashboard", "<ui type> ui". Strong ones: fictional-interface, user-interface-superchannel, interface-bxtmor_fc40, ux-ui-inspiration-ygomnyuw37e
• interaction timing, moving: channels "ui animation" / "micro interactions", e.g. ui-animation-x1j9nqqkflg, digital-ui-micro-interactions

Family: one whole channel, sheeted in full; batch one-liners cite tile numbers.

Pinterest

Web search restricted to pinterest.com. In Claude Code that is WebSearch(query="<role query>", allowed_domains=["pinterest.com"]); a plain site:pinterest.com query returns nothing there. In Codex or another host, use its web search with a domain filter if it has one, else site:pinterest.com.

Resolve pins with pin <url> ...: original media, moving, outbound source (often the Dribbble / Behance / Instagram original), embed (only a thumbnail of a YouTube / Vimeo / reel) and dead.

Primary sites

Follow source from the block or pin first. Otherwise web search restricted to the right home: the studio's site, Behance, Vimeo, a museum's open-access collection or Wikimedia Commons for artworks, the brand's press kit for logos. For like <brand>, the product surface (homepage, docs, app) is read for hex, type and one real component; the images themselves come from are.na channels named for the brand (see "The fence"). Screenshots of docs or settings pages are facts, not references. Dribbble and Behance block scripted fetches (empty 202): cite the original page, take the file through the block or pin.

Assets outside Sketchfab: Poly Haven (CC0, props and textures), the type foundry or Google Fonts, the brand's own SVG, Wikimedia Commons, unsplash for placeholders, GitHub code search (filename:<thing> extension:glb, check the licence).

3D models (Sketchfab)

Use for anything that looks fake when invented in code: anatomy, animals, characters, vehicles, products with recognisable detail, mechanical assemblies, sculptures, interiors. The hand-coded brain was "very bad"; a real GLB fixed it. Keep procedural geometry for orbs, blobs, primitives, extruded logos, lathe objects, noise fields.

find_refs.py sketchfab "<object> [style word]" .context/refs/<topic>/3d [n] [max_faces]# rows: n, name, author, license, faces, animations, glb_mb, textures, tex_max, free_download, page# tiles in .context/refs/<topic>/3d/sketchfab-sheet.png, row-major = nfind_refs.py sketchfab-get <uid> .context/refs/<topic>/3d/<name>.glbnpx -y @gltf-transform/cli inspect <file>    # meshes, animations, bounds

• Search broad. Sketchfab matches every word: R46 subway car → 0, nyc subway car → 24. Object plus one style or era word, then narrow by eye.
• Downloadable only (the script already filters). Prefer CC0 / CC-BY and record the author; flag NC licences in the fill-ins.
• Face count fits the use. A 93k-tri brain was grey mush at 620 px until simplified to ~6k. Check glb_mb and tex_max against the 2 MiB script budget when embedding.
• Parts must match the motion. A lid, arm or door that moves must be its own mesh; say which parts move in the take, confirm with inspect after download.
• Whole object, not a scan fragment or fan variant. Pick from the sheet.
• Download routes, tried in order by sketchfab-get: the Objaverse mirror (allenai/objaverse on Hugging Face, 2022 snapshot, ~800k models, no login, rigs survive; free_download: true marks these, so prefer them when candidates are close); the Sketchfab API with SKETCHFAB_TOKEN; else it prints the page, which goes in assets: with [? download].
• Role. Normally an asset, used as-is with styling on top. A copy reference only when the comp redraws it in another medium. Say which.

Vetting, every time

Matching titles are not references. Pinterest's top sticker peel animation pins were tutorial thumbnails.

1. Download each candidate to .context/refs/<topic>/ with get, named <role-letter><n>.<ext> (m1.mp4, a1.jpg). get fixes the extension to the real type (a png served as .jpg is saved as a1.png); cite the saved path.
2. Look at it. Stills: Read the file. Moving: sheet <file> <out.png> makes a 3x3 contact sheet with timestamps and duration; sheet <file> <out.png> 16 (or 25) when the clip is long and a beat's time range has to be read to ~0.3 s.
3. Keep only if the role's quality is visible: motion, the mechanic happens on screen as more than a cut or fade and it is not a tutorial; art direction, the palette or texture fills most of a 600px+ frame; copy, the real thing, not a remake or recolour; ui, a credible interface at a readable size.
4. Five images and three moving references per brief, plus copy / ui / assets as the beats need them. Runners-up go in .context/refs/<topic>/notes.md, not the prompt.
5. Nothing passes: say so in the fill-ins with the query tried, describe the mechanism in words, and leave the [? motion ref] / [? visual ref] marker in the prompt. Never pass on an unvetted hit to fill a slot.
6. Prefer the striking tile over the matching tile. When two candidates pass, keep the one whose move or look you would want to see regardless of the draft. Passing includes the fence: a tile outside a named style or an attachment's look does not pass, however striking.
7. A film frame, not a web page (SKILL.md rules): a composition reference that is a screenshot at browser scale is a fact for a1's take, not a4.

Structuring the brief from the references

1. Beats from the motion references. Read each contact sheet: what enters, the main action, each transition, the exit, with times from frames_at_s. Assign each clip its beat (together they cover the mechanic, the turn and the camera or exit; number them m1-m3 in beat order), swap the subject, scale to the target duration and state the ratio (m1 is 3.4s, ours ~5s, so 1.5x slower). Name the time range taken (m1 0.0-1.6s).
2. First line and surfaces from the five images. Name the aesthetic in the first line. Then place each image on a beat with the exact quality taken: a3's wet stone is the canvas surface from beat 1, a5's bead row is the end state of beat 2, a2's palette for the ink only. Never "like a1".
3. Subject from the copy reference and assets, on the beat where it first appears, with its fidelity word.
4. The turn and the ideas: SKILL.md steps 4-5. If no motion reference has a turn beat, add one and find its clip or detail reference.
5. Detail references inline on the one beat they affect, scoped.

The refs: / assets: block follows the brief template in SKILL.md.