insta-spot-search

insta-spot-search answers the question those promotional reels refuse to answer — *where is this?* The genre is familiar: a drone shot of a beach or a cafe interior, and a caption saying "follow me and drop a comment, I'll DM you the location." The skill skips the follow-for-location transaction and recovers the place from the clues the video already leaks. The pipeline starts deterministically: yt-dlp pulls metadata, the post caption, comments, and any location tag, then ffmpeg extracts frames (24 at 1024px by default, tuned for reading burned-in captions and signage). Claude reads those frames and builds a clue inventory — signs, bridges, lighthouses, coastline shape, hashtags. Search legs then run in parallel: cross-platform uploader lookup, landmark-signature search (Naver blogs for Korea, travel blogs and TripAdvisor overseas), and geocoding through Kakao Local domestically or Nominatim abroad. A region heuristic (water color, phone area codes, driving side, foreign-language signage) narrows the search area before any query fires. What makes it converge is the verification stance: every candidate is actively attacked before it is accepted, and the final report carries a confidence level (confirmed / likely / candidate), the evidence behind it, and a map link. In one real run a 19-second drone reel was pinned to Sacheonjin Beach in Gangneung by cross-matching an arch footbridge, a yellow breakwater lighthouse, and the granite tide-pool rocks. Accuracy extras are opt-in rather than always-on: zoom refine re-extracts a segment at 2048px to read small signs without re-downloading, narration transcription runs only when `--audio` is passed *and* a Whisper key exists, and a profile scan uses the uploader's recent location tags as a regional prior. Video and frames never leave the local workspace. The guardrail is explicit — the skill serves public places promoted in content, and refuses requests aimed at tracking a private individual's home or movements.