Update to vidIQ

{
  "description": "Mine YouTube comments for what viewers ask, want, and object to, with identities removed, then turn it into hooks, titles, FAQs, or script angles, and draft replies. Use when asked 'what do my viewers want', 'summarize or analyze the comments on this video', 'what is my niche's audience asking for', 'does this idea resonate', or 'reply to this comment'.",
  "included_files": [
    {
      "relative_path": "agents/openai.yaml",
      "size_in_bytes": 448
    },
    {
      "relative_path": "references/discovery-evidence.md",
      "size_in_bytes": 1213
    },
    {
      "relative_path": "references/job-lifecycle.md",
      "size_in_bytes": 724
    },
    {
      "relative_path": "references/live-surface-notes.md",
      "size_in_bytes": 2014
    }
  ],
  "name": "vidiq-comment-insights",
  "skill_md_contents": "---\nname: vidiq-comment-insights\ndescription: Mine YouTube comments for what viewers ask, want, and object to, with identities removed, then turn it into hooks, titles, FAQs, or script angles, and draft replies. Use when asked 'what do my viewers want', 'summarize or analyze the comments on this video', 'what is my niche's audience asking for', 'does this idea resonate', or 'reply to this comment'.\n---\n\n# vidIQ Comment Insights\n\nRead [Live surface notes](references/live-surface-notes.md) first, even when using only supplied evidence.\n\nExtract reusable audience language from comments while protecting commenter identity. Works on\nown or public videos and never reports usernames.\nKeep viewer wording distinct from the creator's transcript and from original copy. Sample named\nvideos, research a whole niche, or draft replies, as the request requires.\n\n## Sample the requested audience\n\n1. Establish topic, intended use, language/market, and period; use a bounded sample.\n   Use supplied videos. Discover additional sources only when needed and within scope.\n2. For discovery, use `vidiq_channel_videos`, `vidiq_youtube_search`, or `vidiq_outliers` to\n   include typical, recent, and breakout examples. Catalog calls require matching\n   `videoFormat: long`, `short`, or `live`; use `popular: false` for recent uploads.\n   YouTube search needs `type: [\"video\"]` so channels/playlists do not enter comment calls.\n   Keep language/format explicit where supported; outliers cannot discover non-English Shorts.\n3. Read [Discovery evidence](references/discovery-evidence.md) when discovering source videos or\n   asserting a recurring pattern. Inspect thumbnails only\n   if visual context changes source selection; missing image access does not invalidate comment\n   analysis. Fetch independent samples in parallel within scope.\n4. Start `vidiq_video_comments` with `videoId`, explicit `order`, and singular `maxResult`\n   (1–100 threads per page). Report the sample cap and selection bias. Count actual returned\n   replies; `replyCount` is not the number inspected. When the live tool returns `nextPageToken`,\n   continue within the sample limit using `pageToken` and the same source, `order`, and\n   `minLikes`. Like filtering can leave a page short or empty while more pages remain.\n   Filter dates locally; undated comments have unknown recency.\n5. Use representative `vidiq_video_transcript` results only when useful; `transcription` is\n   untimed creator language, not viewer evidence.\n\n## Research a niche instead of named videos\n\nFor \"what does my niche's audience keep asking for\" or \"does this idea resonate\", use\n`vidiq_comment_insights` instead of sampling videos one by one. `mode: \"discover\"` with `niche`\n(and optional `language`) returns recurring requests and pain points across a niche-wide cohort;\n`mode: \"validate\"` with `niche` and `idea` tests one proposed idea against comment evidence;\n`mode: \"raw\"` returns underlying threads and is the only mode that accepts `includeReplies`,\n`intents`, `minLikes`, `sort`, `pageSize`, `maxThreadsPerChannel`, and `cursor`, which other modes\nreject. Continue a raw search with `mode: \"raw\"` and `cursor` only, and omit\n`commentPublishedAfter` unless the creator asks for a time window. Every mode returns an\n`mcpJobId`: read [Job lifecycle](references/job-lifecycle.md), then poll `vidiq_job_poll` until\ncompleted. The tool picks its own cohort, so it does not replace `vidiq_video_comments` for named\nvideos or channels, and the Claude.ai directory connector does not expose it; sample videos there.\n\n## Deidentify and count accurately\n\nVerify the video's owner channel from metadata, using `vidiq_get_videos_by_ids` only when needed.\nCompare `authorChannelUrl` in the same identifier form: `/channel/ID` against the owner ID,\nor `/@handle` against a verified owner handle. If the handle is missing, one\n`vidiq_channel_search` with `channelTitleMatch: \"exact\"` can resolve it; require the returned\n`channelId` to match the video's owner before trusting its handle.\n\nAn ID/handle mismatch or display name does not establish non-owner status. Without a comparable\nowner identity, all affected authors remain unverified: paraphrase separately, exclude them from\nverified viewer counts, and use no verbatim fragments. Exclude matched owner comments and replies.\nUse author IDs and handles transiently, never in the language bank or evidence trail.\n\nDecode entities, strip HTML, URLs and handles, and normalize whitespace. Drop sensitive or\nidentifying fragments instead of trying to salvage them by partial redaction. Prefer paraphrases.\nFor useful verbatim evidence, take one contiguous deidentified fragment of at most eight\nwhitespace-delimited words, shorter than the full cleaned comment. Otherwise paraphrase or omit.\nNever stitch fragments or reproduce full comments, even short ones.\n\nCluster questions, desired outcomes, objections, misconceptions, and emotional language.\nCount support by videos: N of M with usable sampled comments, with M out of all selected videos.\nMultiple replies in one thread are not independent video support. Preserve per-video caps,\nreturned/retained threads and replies, exclusions, and date coverage. Empty, disabled, or missing\ncomments are coverage gaps, not absence of audience concerns.\n\nUse `vidiq_keyword_research` only to test a requested or decision-relevant search-demand\nhypothesis; conversational phrasing does not establish search intent. Do not infer demographics\nor sensitive traits from wording.\n\n## Draft replies\n\nFor reply ideas to one specific comment, use `vidiq_generate_comment_replies` with the video\n`title`, the exact `comment`, and optional `tones` (3 to 10 labels; the defaults are General,\nHumorous, Thankful, Witty, and Informal). It returns drafts only; the creator posts the reply they\nchoose on YouTube. The Claude.ai directory connector does not expose this tool.\n\n## Deliver\n\nReturn the requested language bank or audience findings with meaning, support, recency, source-video\nlabels, limitations, and suggested use. Label paraphrases, verbatim evidence, creator language,\nand original copy distinctly.\n\nAdd hooks or content briefs only when requested, using the requested count. Keep all examples\noriginal and the evidence trail deidentified.\n"
}