Update to TubeGrade

{
  "description": "Write a researched YouTube script — study how the top-performing videos on the topic hook, structure and pace it (from their transcripts), then write an original script with timestamped source citations and flagged claims to verify. Use for \"write a script about X\", \"research and script this video idea\".",
  "included_files": [
    {
      "relative_path": "agents/openai.yaml",
      "size_in_bytes": 445
    },
    {
      "relative_path": "references/tubegrade-tools.md",
      "size_in_bytes": 10860
    }
  ],
  "name": "tubegrade-script-writer",
  "skill_md_contents": "---\nname: tubegrade-script-writer\ndescription: Write a researched YouTube script — study how the top-performing videos on the topic hook, structure and pace it (from their transcripts), then write an original script with timestamped source citations and flagged claims to verify. Use for \"write a script about X\", \"research and script this video idea\".\n---\n\n# Research-backed script writer\n\nRead `references/tubegrade-tools.md` first.\n\nGoal: an original, ready-to-record script whose structure is learned from\nvideos that actually over-performed, with every borrowed fact traceable.\n\n## 1. Scope\nAsk only what you don't have: topic, target length (minutes), audience,\nchannel voice (a handle lets you check their style), long-form or Short.\n\n## 2. Find the reference videos — ≈ 15–25 credits\n- `tubegrade_find_outliers` on the topic (`days` 365, `min_ratio` 2), and/or\n  `tubegrade_search_videos` (`upload_date` year, `video_type` long) keeping\n  results with `outlierRatio` ≥ 2.\n- Shortlist the 3 most relevant strong performers (on-topic beats biggest).\n  Tell the user which three you picked and why.\n\n## 3. Study them — 30 credits (hard cap: 4 transcripts without asking)\n`tubegrade_get_transcript` for each of the 3. If one is unusable (no speech,\nmusic only, wrong topic), replace it once — a 4th transcript. Anything beyond\nthat, ask first. From each usable transcript, extract:\n- the **hook** (first ~30–45 seconds): how it earns attention\n- the **structure**: sections with timestamps, and where the payoff lands\n- **retention devices**: open loops, previews, pattern breaks, stakes\n- **facts and claims** worth using, with their `[MM:SS]` timestamp\nOptional: `tubegrade_get_comments` (5) on the strongest video to catch what\nviewers felt was missing — fill that gap in the new script.\n\n## 4. Write\n- Original wording throughout. Learn the structure; never paraphrase a\n  transcript line by line.\n- Open with a hook that states the promise and the stakes within ~20 seconds.\n- Sections with on-screen/B-roll notes in [brackets].\n- After any fact, technique or statistic taken from a reference, add a\n  citation line: `Source: <video title> — <link> at [MM:SS]`.\n- Mark anything that needs outside verification (health, money, legal,\n  statistics) with **[VERIFY]** — transcripts are not proof.\n- Match the requested length (~140–160 spoken words per minute).\n\n## 5. Deliver\n1. The 3 reference videos (link, views, outlier ratio) and what each taught.\n2. The script.\n3. A **[VERIFY] checklist** of claims to confirm before recording.\n4. Offer packaging next → **tubegrade-packaging-studio**.\n\n## Don't\n- Don't present a reference video's script, jokes, or unique stories as the\n  user's.\n- Don't state medical, financial, or legal claims as settled — flag them.\n"
}