When to Use

The user shares a YouTube link (watch, youtu.be, shorts, live) and wants to know what it says, whether it is right, or a report about it. One video per run; for a playlist, ask which videos or run once per video.

Procedure

1. Fetch. skill_run(name="youtube-analysis", args="<url>") (add --lang en if the user writes English). It answers JSON with the folder (reports/youtube/<id>/), the video facts, chapters, the transcript kind and the contact sheets. status: blocked → tell the user the what and stop. status: error → read error.
2. Watch. This is the whole point; do not skip it.
   • read_file transcript.md to the end (use offset for long ones). It is the soundtrack, with [mm:ss].
   • read_file every sheets/sheet-NN.jpg: 4×4 stills with times, the picture track. Note charts, on-screen text, demonstrations, who is on screen, sponsor segments.
   • read_file source.json for the description, tags and chapters (links in the description often name the sources the video relies on).
   • Without a transcript (kind: none) work from the stills, description and chapters, and say so in the report.
   • Auto captions (kind: auto) mishear names and numbers: confirm them from the stills or the description.
3. Classify. Read references/categories.md, pick one of the 16 categories and answer its questions.
4. Fact-check if the category says so: references/fact-check.md. Use web_search and web_fetch (the web-research skill's habits apply: open the page, two sources for anything important, note dates).
5. Cover. Write <folder>/cover.json:

   {"title": "<the report's title, the user's language, ≤ 40 characters>",
    "tagline": "<one sentence: who says what>",
    "category": "<key from categories.md>",
    "points": ["<takeaway 1, ≤ 26 characters>", "<takeaway 2>", "<takeaway 3>"],
    "verdict": {"label": "<整體判定, e.g. 大致正確>", "tone": "ok | warn | err | info"}}
   

   Leave out verdict for categories that are not fact-checked. Then skill_run(name="youtube-analysis", args="cover <folder>") and look at cover.png with read_file; shorten the title or points if anything is cut with "…".
6. Stills for the report. The fetch saved stills in frames/ (named tMMSS.jpg). For a specific moment: skill_run(name="youtube-analysis", args="frame <folder> 12:30").
7. Write <folder>/report.md from references/report-template.md.
8. Check. skill_run(name="youtube-analysis", args="check <folder>"). Fix every listed problem and check again until status: ok.
9. Deliver in chat: the cover as an image, a link to the report, the one-sentence summary and the overall verdict, e.g.

   ![封面](reports/youtube/<id>/cover.png)
   **[開啟完整報告](reports/youtube/<id>/report.md)**　一句話摘要… 整體判定：大致正確
   

   SinoBee opens the linked report in its document viewer (cover, tables and stills rendered). The folder travels as a whole: cover.png and frames/ are referenced by relative paths.

Rules

• Watch before writing: every statement in the report comes from the transcript, the stills, the description or a fetched source, never from the title alone or from memory of the topic.
• Timestamps everywhere: outline, analysis and fact-check rows link to https://www.youtube.com/watch?v=ID&t=<s>s.
• Separate what the video says from your assessment; separate facts from opinions.
• Quote only short phrases from the video (lyrics and scripts especially); summarise the rest.
• Health and finance reports end with a one-line disclaimer (see categories.md).
• Use skill_run for the scripts; do not call yt-dlp or the scripts through shell.

Gotchas

• HTTP 429 or "Sign in to confirm you're not a bot": YouTube is limiting this network. The script reports blocked; try again later or from another network. Don't loop.
• Members-only, private and age-restricted videos cannot be fetched without the user's cookies; say so.
• Live streams still in progress have no final transcript; analyse the replay later.
• Very long videos (over 2 hours): read the transcript in parts and keep notes per chapter in working memory before writing.
• Stills are YouTube storyboard frames (about 320×180); good for seeing what is shown, too small for fine print.

Verification

check returns status: ok; the cover shows the right category, title and verdict; every fact-check row has a verdict and a source (except 無法查證 / 觀點); the report opens with the cover and reads well for someone who has not seen the video.