Update to vidIQ

{
  "name": "vidiq-retention-analysis",
  "description": "Analyze an authorized YouTube video's retention to investigate drop-offs, pacing, expectations, or structure. Align findings with timestamped scenes only when visual analysis is within scope and available.",
  "included_files": [
    {
      "relative_path": "agents/openai.yaml",
      "size_in_bytes": 416
    },
    {
      "relative_path": "references/job-lifecycle.md",
      "size_in_bytes": 724
    },
    {
      "relative_path": "references/live-surface-notes.md",
      "size_in_bytes": 2014
    }
  ],
  "skill_md_contents": "---\nname: vidiq-retention-analysis\ndescription: Analyze an authorized YouTube video's retention to investigate drop-offs, pacing, expectations, or structure. Align findings with timestamped scenes only when visual analysis is within scope and available.\n---\n\n# vidIQ Retention Analysis\n\nRead [Live surface notes](references/live-surface-notes.md) first, even when using only supplied evidence.\n\nInvestigate retention changes with private analytics and relevant content evidence.\nTreat explanations as hypotheses, not proof of why every viewer left.\n\n## Establish authorized evidence\n\n1. Identify video, format, and review dates; keep analysis and optional watch/poll work within scope.\n2. Resolve authorized IDs with `vidiq_user_channels`. If no video is supplied, discover candidates\n   with `vidiq_channel_videos`, `popular: false`, and confirmed `videoFormat: long`, `short`,\n   or `live`; confirm the candidate. Use `vidiq_get_videos_by_ids` and require its `channelId`\n   to match an authorized ID before private analytics. Otherwise offer a public-content critique\n   with retention evidence unavailable.\n3. Request `vidiq_channel_analytics` with explicit dates, `report: \"audience_retention\"`,\n   `filters: \"video==VIDEO_ID\"`, `metrics: [\"audienceWatchRatio\", \"relativeRetentionPerformance\"]`,\n   and `maxResults: 100`. The equivalent `dimensions: [\"elapsedVideoTimeRatio\"]` still requires\n   that video filter. Read `columnHeaders` before interpreting the 100-point curve.\n\n## Interpret the curve\n\n`audienceWatchRatio` is segment watches relative to video views; repeat viewing can exceed 1\n(100%). `relativeRetentionPerformance` compares videos of similar length on a 0–1 scale:\n0.5 is the median, with higher values above it and lower values below it. This is comparative\nrank, not the percentage of this video's viewers; see [YouTube's metric definition](https://developers.google.com/youtube/analytics/metrics#relativeRetentionPerformance).\nIf requested, `totalSegmentImpressions` counts segment watches, not unique viewers or thumbnail\nimpressions. Report available sample/denominator context; scope any extra read needed for missing counts.\n\nIdentify material drops, recoveries, or sustained changes against surrounding observations and\nrelative-retention context. Use actual non-null endpoints; percentage-point change is 100 times\nthe ratio difference. Record endpoints, interval, magnitude, and why it warrants attention.\nDo not impose universal editing cutoffs or interpolate gaps into evidence. Sparse curves, small/unknown\nsamples, and near-end drops lower confidence without establishing an editing fault.\n\n## Align content only with appropriate evidence\n\n`vidiq_video_transcript` returns untimed `transcription`. Use it for spoken themes.\nFor scene alignment within the requested analysis, verify the source before watching:\n`vidiq_video_watch` serves long-form; `vidiq_watch_shortform_content` serves Shorts.\nVerify eligibility separately for live streams or recordings. Before a watch call, read\n[Job lifecycle](references/job-lifecycle.md) for submission, bounded polling, and recovery.\n\nElapsed ratio times known duration gives an approximate analytics position, not an observed\nscene timestamp. Attach scenes or timed hypotheses only when completed watch evidence supports\nthe alignment; scene ranges are not frame-accurate ground truth. Never attach a transcript quote\nto a time range without evidence at that precision.\n\nWithout watch evidence, separate numeric zones from untimed content hypotheses. Use\n`vidiq_video_comments` only for useful corroboration, preferring deidentified paraphrases;\ncomments are not representative of every viewer.\n\n## Deliver\n\nReport the scoped retention findings with actual metric values/deltas, approximate analytics\npositions, watched scenes when available, hypotheses, confidence, and material sample limits.\nInclude requested watch job status and distinguish missing visual evidence.\n\nWhen the creator requests editing advice, prioritize supported structural changes and the next\nexperiment; preserve what worked. Rewrite the opening or section order only when requested.\n"
}