Museum Art Agent Skill

museum-art: Public-Domain Artwork for Visuals

> Standing rule (adopted 2026-07-24, from Eric Li's post on museum open-access): whenever a visual needs a real image with aesthetic weight, source public-domain museum artwork first - over AI-generated imagery and over generic stock. Curated, historically significant art reads as credible and sophisticated; AI-gen reads as slop. This stacks with, and reinforces, the no-ai-slop skill and your house image style. It does NOT replace the generative editorial-illustrations skill (that owns claim-driven diagrams/figures) or your house chart style - use museum art for photographic/hero/decorative/mood imagery, generative figures for data and concept diagrams.

When to reach for this

• Blog post hero images, section breaks, mood imagery (the blog-publish image step).
• Deck/slide backgrounds and section dividers, social cards, essay figures, spec cover art.
• Any time the instinct is "generate an image" for something decorative or evocative. Stop and pull a real painting instead.
• NOT for: product screenshots, UI mockups, data charts, logos, or claim-driven explanatory diagrams (those are editorial-illustrations / real captures).

Decision: which source

• Met -> Cleveland -> SMK first. All keyless, one JSON hop, CC0/PD, broad collections. Fastest path to a hi-res image.
• Need Dutch/Flemish masters or decorative arts? Rijksmuseum (keyless, 3 hops).
• Need European antiquities/photography and keyword isn't essential? Getty (keyless, SPARQL).
• Nothing fits, or you want a cross-museum search? Wikimedia Commons API (keyless aggregator, normalized license metadata) is the best general fallback.
• "Historical illustration/engraving/old photo" rather than fine-art painting? Go straight to Internet Archive or Wikimedia Commons.
• Science/health/medical essay figure? Wellcome Collection (keyless, CC0/PD).

How to fetch in THIS environment

• Use WebFetch to hit the JSON search endpoint, then WebFetch/download the returned image URL. These APIs are server-side reachable; no browser needed.
• Exception - Art Institute of Chicago images: the JSON API (api.artic.edu) is fine, but the image host www.artic.edu/iiif/... 403s scripted/curl fetches via a Cloudflare bot challenge. Use the JSON metadata from AIC, but download the actual image through a real/headless browser (browser-harness) or prefer a different museum for the image bytes.
• Always filter for public domain in the query AND spot-check the per-image license flag before shipping (see Licensing).

Freshness over caching (mandatory)

Fetch fresh per need. Do NOT build a reusable local pool of downloaded images to draw from. A small cached set gets reused everywhere and becomes the new "same stock photo on every post" - sameness is a form of slop, and the variety of a huge open collection is the entire point. Fetching is keyless and sub-second, so there is no cost reason to cache pixels.

• Cache recipes/metadata, not images - that is what this skill's references/ already are.
• Commit an image only into the specific artifact that uses it (a post's assets/, a deck's media) once chosen - for provenance and offline builds. That is an artifact asset, never a shared library other artifacts pull from.
• Each new visual = a fresh query. Vary the search terms and the source museum so consecutive posts do not converge on the same few crowd-pleasers.

Verified keyless recipes (2026-07-24, all live-tested)

1. The Met - best all-around default

• Base: https://collectionapi.metmuseum.org/public/collection/v1/ - no key, 80 req/s, CC0 where isPublicDomain:true.
• Search: GET /search?q=<term>&hasImages=true&isPublicDomain=true -> {total, objectIDs[]}
• Object: GET /objects/{id} -> read primaryImage (full-res JPEG, static, no IIIF hop) or primaryImageSmall (web-large).
• Example: https://collectionapi.metmuseum.org/public/collection/v1/search?q=sunflowers&hasImages=true&isPublicDomain=true

2. Cleveland Museum of Art - only one with archival TIFF

• Base: https://openaccess-api.clevelandart.org/api/artworks/ - no key, CC0.
• Search: GET /api/artworks/?cc0=1&has_image=1&limit=10 (add &q=monet, &skip=10).
• Image fields on each result: images.web.url (900px), images.print.url (3400px JPEG), images.full.url (archival TIFF). Directly downloadable from openaccess-cdn.clevelandart.org.
• Confirm share_license_status == "CC0" per result.

3. SMK (Denmark) - one-hop, broad European/Nordic

• Base: https://api.smk.dk/api/v1 - no key. License: Public Domain Mark 1.0 (functionally CC0).
• Search: GET /art/search?keys=*&filters=[public_domain:true]&filters=[has_image:true]
• Read image_native directly from each result (no chain).
• Gotcha (verified): pass filters as a repeated query param, one [field:value] bracket each. Concatenating filters=[public_domain:true][has_image:true] returns 200 but silently ignores the second condition.

4. Rijksmuseum - Dutch/Flemish masters (keyless, 3 hops)

• Base: https://data.rijksmuseum.nl/search/collection (Linked Art, no key).
• GET /search/collection?type=painting&imageAvailable=true -> walk object -> VisualItem -> DigitalObject -> access_point[0].id is a ready IIIF URL.
• Mixed license: mostly CC0/PD but some CC BY 4.0 - check the per-object rights block (a CC BY item requires attribution).

5. NGA (Washington) - bulk/offline, no live search

• CSV: https://raw.githubusercontent.com/NationalGalleryOfArt/opendata/main/data/published_images.csv - filter openaccess=1.
• Image (IIIF): https://api.nga.gov/iiif/{uuid}/full/full/0/default.jpg
• Per-image openaccess=0 rows are NOT open (resolution-capped, rights-restricted). Only openaccess=1 is free.

6. Art Institute of Chicago - Impressionism/European (image host caveat)

• JSON: GET https://api.artic.edu/api/v1/artworks/search?query[term][is_public_domain]=true&fields=id,title,image_id - no key, CC0.
• Image: build https://www.artic.edu/iiif/2/{image_id}/full/843,/0/default.jpg - but Cloudflare 403s scripted fetches; download via browser or use another museum for bytes.

7. Getty - European painting/photography/antiquities (SPARQL)

• Base: https://data.getty.edu/museum/collection/ - no key. No keyword search; use SPARQL with ?obj crm:P138i_has_representation ?img which returns a ready IIIF image URL in the same query.
• Dataset is CC0; per-image rights inconsistently populated - cross-check the object's public getty.edu page (a "Download" button flags true Open Content) before commercial use.

Key-gated + aggregator sources (know these too)

• Smithsonian (needs free api.data.gov key; DEMO_KEY works at 30 req/hr): GET https://api.si.edu/openaccess/api/v1.0/search?q=<term> AND online_media_type:Images&api_key=KEY; CC0; check media[].usage.access=="CC0" per image. Extremely broad (SAAM, NPG portraits, Freer/Sackler Asian, Cooper Hewitt design).
• Wikimedia Commons (keyless, best aggregator fallback): https://commons.wikimedia.org/w/api.php?action=query&generator=search&gsrsearch=<term>&gsrnamespace=6&prop=imageinfo&iiprop=url|extmetadata; license per file in extmetadata.LicenseShortName; raw bytes via Special:FilePath/{filename}.
• Wellcome Collection (keyless): https://api.wellcomecollection.org/catalogue/v2/works?query=<term>; CC0/PD medical/scientific/historical imagery.
• Internet Archive (keyless): https://archive.org/advancedsearch.php?q=<term>&fl[]=identifier&output=json; PD book illustrations, engravings, historical photos, ephemera.
• Europeana (free key): https://api.europeana.eu/record/v2/search.json?wskey=<KEY>&query=<term>; 50M+ items across European institutions.
• Harvard Art Museums (free key): https://api.harvardartmuseums.org/object?apikey=<KEY>&q=<term>.
• Yale LUX (keyless, Linked Art like Getty): https://lux.collections.yale.edu/api/search/...; strong British art + rare books.
• Paris Musees (keyless, explicit CC0): parismuseescollections.paris.fr / opendata.paris.fr; French painting/decorative arts.
• NYPL Digital Collections (free key): https://api.repository.library.nyc/...; PD prints, maps, photos, illustrations.
• DPLA (free key): https://api.dp.la/v2/items?q=<term>&api_key=<KEY>; federated US collections.

Licensing rules (apply before shipping any image)

• True CC0 (no attribution required): Met, Cleveland, Art Institute of Chicago, Smithsonian, NGA (dataset-level; image gated by per-image openaccess=1).
• Public Domain Mark / open-but-not-CC0 (free to use, credit requested not required): SMK, and the PD subset of Rijksmuseum.
• Verify per-image, do NOT blanket-trust collection-level "open access": Getty, Rijksmuseum (CC BY items require attribution), NGA and Smithsonian per-image flags.
• Rule: when the API exposes a rights/license field, filter on it in the query AND spot-check it on the chosen image. Never assume "the collection is open access" implies "this specific image is."
• Good practice everywhere: a one-line credit ("Digital image courtesy of [Museum]") costs nothing and covers mixed-license collections. When masking/compositing per house style, keep the credit.

Relationship to other skills

• Stacks with no-ai-slop and the house image style: real museum art is the anti-slop default for evocative imagery.
• Complements editorial-illustrations (generative claim-driven diagrams) and dataviz (charts) - those own explanatory graphics; this owns photographic/artwork/mood imagery.
• Feeds blog-publish, social-media-kit, weekly-ai-slide, deck and essay work at their image-sourcing step.

References (full per-museum recipes, verified)

references/_synthesis.md (cheat-sheet) + one file per museum (met.md, cleveland.md, smk.md, rijksmuseum.md, nga.md, artic.md, getty.md, smithsonian.md) with live-tested example URLs, field maps, and gotchas.