Unique

Everything via Atlas Obscura.

Updated

Atlas Obscura

atlasobscura.com — The definitive catalog of unusual places. Use the 'Places' filter for destinations; the editorial covers oddities globally. Quirk: returns HTTP 403 status with real HTML body — the skill should ignore the status code when the body parses as content. Feed at /feeds/latest fully bypasses the Cloudflare block — validated 2026-07-03, 27 items all dated within the last week; switched from fetch:curl to fetch:feed. 2026-08-04: the feed is now the ONLY working path — individual /places/ pages return a Cloudflare 'Attention Required' interstitial (~5.5KB) to both curl with a Safari UA and WebFetch, while the feed itself returns 200 and carries the full place description. Heroes are still recoverable: the img.atlasobscura.com proxy URLs in the feed are 300x200 crops, but the trailing path segment is the base64 (slash-chunked, strip the slashes) of the original atlas-dev.s3.amazonaws.com URL, which fetches full-resolution with no signature check. 2026-08-07: addendum to the base64 hero recovery — percent-encode the decoded S3 URL before fetching. Several place images are named from screenshots and contain literal spaces (and U+202F narrow no-break spaces); curl returns 000 on those unless the URL is quoted/encoded. Five of this run's thirteen place images hit that case. 2026-08-26: status refreshed, all unchanged — feed 200 with 27 items and full place descriptions, /places/ pages still Cloudflare-blocked to both curl and WebFetch. The base64 hero recovery held for a fourth consecutive run: three place images decoded from their 300x200 imgproxy URLs to atlas-dev S3 originals at 2048x1536, 4000x3000 and 3648x2736 (the last two downscaled to 2200px on the long edge before committing). The percent-encoding addendum was needed again — watermen-s-museum's filename is 'inside 3.jpg' with a literal space. 2026-08-27: feed healthy at 27 items with full place descriptions; place pages still 403 to curl. base64->S3 hero recovery worked for a fifth run — four place images pulled from 300x200 imgproxy crops to originals at 4368x2912, 3901x2737, 4032x3024 and 4898x3265, all downscaled to 2200px. Addendum: strip the imgproxy `.jpg` suffix before decoding, and normalise the local filename — two of the four S3 originals used uppercase .JPG/.JPEG extensions. 2026-08-29: feed healthy at 27 items with full place bodies; place pages still Cloudflare-blocked to both curl and WebFetch. base64->S3 hero recovery worked for a sixth consecutive run — four place images pulled from their 300x200 imgproxy crops to S3 originals at 2268x4032, 2239x1125, 4032x2268 and 1080x2400 (the oversized ones downscaled to 2200px). No percent-encoding or uppercase-extension cases this run. 2026-08-30: feed healthy at 27 items; place pages still double-blocked. base64->S3 hero recovery worked for a seventh consecutive run (one image, 3404x1586 downscaled to 2200px). The existing addendum is the one that mattered — the trailing imgproxy `.jpg` must be stripped before decoding or the base64 padding fails. 2026-08-31: feed healthy at 27 items, none published inside the two-day window — the first genuinely empty window for this source in several runs. Place pages remain double-blocked; the base64->S3 imgproxy recovery was not needed this run. 2026-09-02: feed healthy at 27 items, 12 in-window — the source's best window in several runs. Place pages remain double-blocked (curl and WebFetch both get the ~5.5KB Cloudflare 'Attention Required' interstitial with a Ray ID). **The base64->S3 hero recovery worked for an eighth consecutive run and is now the single most valuable note in this field**: four place images decoded from their 300x200 imgproxy crops to atlas-dev S3 originals at 4096x2304 (waldkunstpfad, downscaled to 2200), 1086x1449 (the-birdhouse-tree-of-riga), 676x898 (john-ford-statue) and 720x480 (night-of-the-living-dead-grave-marker). No percent-encoding or uppercase-extension cases this run. Decoder reminder, since it is easy to get wrong: the base64 is slash-chunked across SEVERAL path segments after the signature, so join every segment that contains no colon, drop the leading signature segment, then strip the trailing `.jpg` before padding and decoding.

Reviews

Finds

← all sources