{"id":51542,"plugin_id":"plugin_asdk_app_6ac633e11e548191b9789c97543b311d","kind":"skill","collection_source":"plugin_package","comparison_source":null,"observed_at":"2026-10-24T13:46:41.399Z","digest":"47dd5c0e524e60ef9a427b2a9eb0c9f53ba2ddd6f866a7ea1a2e3215ed447780","against":null,"payload":{"description":"Find a vaguely remembered YouTube video by reconstructing clues, refining searches, comparing returned metadata, and asking the user to confirm a candidate. Use when the title, channel, or link is forgotten or uncertain. Keep possible matches tentative until confirmed. Do not use for generic recommendations, an exact-ID display request, watch-history access, or guaranteed identification from an unseen scene.","included_files":[{"relative_path":"agents/openai.yaml","size_in_bytes":266}],"name":"rediscover-a-video","skill_md_contents":"---\nname: rediscover-a-video\ndescription: Find a vaguely remembered YouTube video by reconstructing clues, refining searches, comparing returned metadata, and asking the user to confirm a candidate. Use when the title, channel, or link is forgotten or uncertain. Keep possible matches tentative until confirmed. Do not use for generic recommendations, an exact-ID display request, watch-history access, or guaranteed identification from an unseen scene.\n---\n\n# Rediscover a Video\n\n## Inputs\n\nCollect clues already present in the conversation: subject, remembered title\nwords, creator, language, approximate period, and distinctive details. Separate\nconfident clues from uncertain memories. Ask at most one high-value question\nbefore searching when a missing clue would meaningfully narrow the results.\nClarify whether a remembered date means upload date or when the user watched.\n\n## Workflow\n\n1. If an exact ID or usable YouTube URL is supplied, extract the ID and skip\n   discovery. Otherwise, summarize the strongest clues briefly and form a query\n   combining the topic with distinctive title words or a creator clue.\n2. Call `search-videos` with `{ query, reason? }` and read `results`. Compare\n   returned titles, descriptions, channels, and publication dates with the clues.\n   Treat a remembered scene as a search hint, not as something confirmed by the\n   metadata. Do not assume upload date equals the user's viewing date.\n3. Present at most three plausible candidates with their titles, channels,\n   returned `watchUrl` values, supported matches, and unresolved or conflicting\n   clues. Mark the candidates tentative. Do not claim an exact match from\n   search rank or similarity alone.\n4. Ask which candidate looks familiar or request one discriminating clue. Use\n   the reply to refine the next query: keep strong clues, relax an uncertain\n   one, and avoid repeating the same failed search. Deduplicate IDs across rounds.\n   Limit discovery to three rounds before offering the best remaining shortlist\n   and the specific missing clue that would help continue.\n5. Call `display-video` with `{ videoId, reason? }` for a candidate the user\n   chooses, an unambiguous metadata match offered for verification, or a supplied\n   exact ID when viewing is wanted. Do not display every tentative candidate.\n   Accept only `status: ready` with a video. Explain `invalid_video_id`,\n   `video_not_found`, or a tool error; offer another actual candidate or request\n   a corrected link without declaring the search resolved.\n6. Invite the user to verify by watching. Mark the video confirmed only after\n   they explicitly recognize it. If rejected, record the useful exclusion in\n   the conversation and continue within the search limit or with a new clue.\n\n## Deliverable\n\nProvide either a user-confirmed video with its title, channel, and watch link,\nor a clearly tentative shortlist with matched clues, conflicts, and the most\nuseful next question. State honestly when no plausible match was found.\n\n## Boundaries\n\n- Do not access or claim access to YouTube accounts, viewing history, saved\n  playlists, private widget state, or playback events.\n- Do not rely on tagging, transcripts, or automatic scene recognition. Use the\n  user's remembered details as clues without inventing corroborating content.\n- Do not claim an absent result proves deletion or that all of YouTube has been\n  searched. A limited search cannot establish either conclusion.\n- Use only the registered tool arguments; channel and date words in `query`\n  are search hints, not enforced filters. Keep clue tracking in the conversation.\n- If the user explicitly requests text only or no video display, call neither\n  tool; help reconstruct clues from their supplied material instead.\n"},"changes":[],"summary":"First saved snapshot. No earlier version is available for comparison.","summary_kind":"deterministic","summary_metadata":{}}