Update to TubeGrade

{
  "description": "Mine YouTube comments for what viewers actually ask, want, complain about and praise — in their own words — and turn it into video ideas, title language and script fixes. Use for \"what do viewers want\", \"analyze the comments on this video\", \"find content gaps from comments\".",
  "included_files": [
    {
      "relative_path": "agents/openai.yaml",
      "size_in_bytes": 440
    },
    {
      "relative_path": "references/tubegrade-tools.md",
      "size_in_bytes": 10860
    }
  ],
  "name": "tubegrade-comment-insights",
  "skill_md_contents": "---\nname: tubegrade-comment-insights\ndescription: Mine YouTube comments for what viewers actually ask, want, complain about and praise — in their own words — and turn it into video ideas, title language and script fixes. Use for \"what do viewers want\", \"analyze the comments on this video\", \"find content gaps from comments\".\n---\n\n# Comment insights\n\nRead `references/tubegrade-tools.md` first.\n\nGoal: viewer demand and language, counted honestly from a bounded sample.\n\n## 1. Scope\nPick the sample:\n- **Specific videos** the user names, or\n- **A topic**: find 3–5 relevant, well-performing videos with\n  `tubegrade_find_outliers` or `tubegrade_search_videos` (5–10 credits), or\n- **A channel**: `tubegrade_get_channel_outliers` or `tubegrade_channel_videos`\n  and take its 3–5 most relevant recent videos.\n\n## 2. Collect — ≈ 15–30 credits\n`tubegrade_get_comments` per video (5 each), `limit` 50–100, `sort` top. Add a\n`newest` pull on one video only if recent reactions matter (e.g. a fresh\nupload). Keep the sample to 5 videos unless the user asks for more.\n\nIf the user wants to know what a video *said* that triggered reactions,\n`tubegrade_get_transcript` (10) on that one video.\n\n## 3. Analyse\nRead every comment you pulled. Sort them into:\n- **Questions** — things viewers asked that the video didn't answer\n- **Requests** — \"do a video on…\", \"part 2\", follow-ups\n- **Pain points / objections** — confusion, disagreement, problems\n- **Praise** — what specifically landed (a moment, an explanation, a format)\n- **Language** — the exact phrases viewers use for the problem and outcome\n\nCount how many comments fall into each theme across the sample (approximate\nis fine; say \"about\"). Weigh likes: a highly liked comment speaks for many.\nIgnore spam, self-promotion and pure emoji.\n\n## 4. Deliver\n1. **Sample** — the videos (linked) and how many comments you read.\n2. **Top themes** — ranked, each with an approximate count and 1–2 short,\n   representative quotes (trimmed; no usernames).\n3. **Video ideas** — 3–5 directly from unanswered questions and requests.\n4. **Words to use** — viewer phrases for titles, hooks and thumbnails.\n5. **Fixes** — if these are the user's own videos, what to clarify or add.\n\n## Don't\n- Quote comments without the commenter's name unless the user asks who said\n  it; `isChannelOwner` marks the creator's own replies.\n- Don't generalise beyond the sample — say how many comments and videos it covers.\n"
}