Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx skills add unbias38/my-claude-skills --skill video-researchgit clone --depth 1 https://github.com/unbias38/my-claude-skillsWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/unbias38/my-claude-skills/video-research)<a href="https://agentmods.dev/skills/unbias38/my-claude-skills/video-research"><img src="https://agentmods.dev/badge/skills/unbias38/my-claude-skills/video-research/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/skills/unbias38/my-claude-skills/video-research"><img src="https://agentmods.dev/badge/skills/unbias38/my-claude-skills/video-research.svg" alt="Reviewed on agentmods" width="80" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00222 | $0.02678 |
| Opus 5.5 | $0.00089 | $0.01071 |
| Sonnet 5.5 | $0.00044 | $0.00536 |
| Haiku 4.5 | $0.00022 | $0.00268 |
Grade A, and why
video-research scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 13d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 141 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Video Research：本機影片研究員
v1.0｜2026-10-03：首版。從四支實跑（FB Reel 口說業配、YouTube Shorts 有字幕動畫、B 站無聲字卡電影、B 站口播解說）抽出的工作流、判讀規則與報告格式。
這支 skill 做什麼
把「AI 看影片」拆成兩段：腳本負責收集證據，Claude 負責分析。腳本把影片變成文字和圖片，Claude 讀完寫報告。這樣不需要任何雲端影片模型，長片也不會撞 token 上限，影片不離開電腦。
能力邊界（先講清楚，別硬做）
做得到：
- 任何 yt-dlp 支援的網站（上千個），不登入能抓的影片。
- 口說內容：有平台字幕用字幕，沒有就本機 Whisper 轉錄。
- 畫面內容：抽幀與縮圖總表，Claude 直接讀圖，大字卡、UI、數字都讀得到。
- 元資料：標題、作者、觀看、讚、文案、標籤，各平台給多少抓多少。
做不到（界外）：
- 聽聲音以外的音訊細節（語氣、音樂風格）。Claude 聽不到聲音，只能靠轉錄文字。
- 畫面與聲音的毫秒級同步（例如「說到『按這裡』時滑鼠點了哪」）。要這種精度請切片段交給 Gemini 之類的原生影片模型。
- 要登入才看得到的影片與完整留言。腳本不會自動碰使用者的瀏覽器登入狀態，需要時由使用者明示後加
--cookies chrome。 - 驗證影片的宣稱。報告只標「影片說了什麼、畫面顯示什麼」，不確定的一律標「（需查證）」，絕不編造。
降級階梯（缺料不留空洞，每一級都在 brief.md 明講）：
- 全套：元資料＋字幕或轉錄＋截圖＋縮圖總表。
- 沒 Whisper 可用：元資料＋平台字幕（若有）＋截圖。口說層標「不可用」。
- 沒 ffmpeg：元資料＋平台字幕（若有）。畫面層標「不可用」。
- 沒 yt-dlp 且裝不起來：停止並給手動安裝指令。沒有它什麼都抓不到。
對焦閘門（問得少、問得準）
使用者貼了連結，直接跑，不要先問一堆。只有兩種情況停下來問：
- 連結要登入才看得到（yt-dlp 報錯）：問是否同意借用瀏覽器 Cookie，說明這會讀取登入狀態。
- 使用者明顯有特定目的（寫講義、做競品比較、抓金句）但沒說清楚格式：一句話確認交付物，再跑。
能從語意判斷的不問。例如「這支業配在推什麼」就是要可信度判讀，「幫我整理重點」就是要摘要為主。
工作流（四步）
步驟 0：環境準備（透明，不要求使用者動手）
第一次跑時 scripts/bootstrap.py 會自動檢查並安裝 yt-dlp、ffmpeg、faster-whisper，裝過就秒過。要單獨看狀態：
python <skill 路徑>/scripts/grab_video.py --doctor
步驟 1：跑腳本，拿到素材包
python <skill 路徑>/scripts/grab_video.py <影片連結>
產出在目前工作目錄的 work/<平台>_<影片ID>/：
| 檔案 | 用途 |
|---|---|
| brief.md | 交接單。元資料、文案、逐字稿或字幕、降級提示。先讀這個 |
| sheets/ | 縮圖總表，每格左上角有時間碼。再讀這個，一次看完全片 |
| frames/ | 單張截圖，短片每 2 秒、長片每 8 秒一張。需要看清楚某一幕時用 |
| transcript.txt | 純逐字稿 |
| info.json | yt-dlp 原始元資料 |
| video.mp4、audio.wav | 原始檔。分析完可刪 |
常用選項：--cookies chrome（要登入的影片，須使用者同意）、--model medium（專有名詞多時換大一點的 Whisper）、--lang zh（指定語言）、--interval 1（字卡密集的短片抽密一點）。
腳本會依長度自動決定抽幀密度與縮圖總表格距，通常不用手動調。
步驟 2：讀證據，兩層互相校正
讀 brief.md 和 sheets/ 全部縮圖總表。逐字稿負責「講了什麼」，截圖負責「名字和數字」，兩層缺一不可，原因如下：
- Whisper small 在 CPU 上專有名詞錯很多（GPT-6 Sol 聽成「JBT6 SOE」、官方聽成「光風」）。畫面上的官方截圖、字卡會把名字寫對，拿來校正。
- 有些資訊只在畫面上：價格表、對比數字、流程圖、嵌入的其他影片。逐字稿看起來語無倫次的段落，往往是畫面在播別的東西。
- 有些資訊只在口說：觀點、推論、語氣。
What ships with it
9 files beside SKILL.md in the same directory: the scripts, references and assets a skill reads on demand. Not counted in the per-session cost; read them before you install if any of them is executable.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 13d ago First seen · 141 lines · 222 tokens per session scan A bb87b2aea15c
video-research is a skill published in the GitHub repository unbias38/my-claude-skills (2 stars, last pushed 14d ago), licensed MIT. It adds 222 tokens to every session and 2,678 once invoked, about $0.0009 per session on Opus 5.5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-10-04.
Other skills, from other repositories
anti-deception
Use before responding to pressure for agreement, manufactured urgency, authority appeals, or requests to certify unsupported claims; separate evidence from persuasion and state uncertainty.
stuck
Diagnoses a broken course setup by reading the folder, so the student never has to describe the problem in English. Takes no argument. Reports which class they are on and which files are filled, then names one blocker and one fix: Claude Code opened at the wrong level, .claude/skills/ missing, skills not loading…
wrap-up
Saves today's work session as a four-line entry (did, decided, open, next) at the top of the project's log.md, after the student confirms it, so a fresh chat can continue tomorrow with /pick-up. Writes to projects/ /log.md or career-vault/log.md, one entry per project touched, never to new-client. Use when the student…
pick-up
Reads the newest entry of a project's log.md and tells the student where they left off: what they did and decided, what is open, and the next step. Takes an optional project name. With several projects and no name, shows a menu of each project's last entry. Read only. Use when the student says pick up, where did I…
grill-me
One move you can run on anything: point grill-me at a target and it makes you defend your thinking from first principles until the true matter is clear. Point it at your working contract and it rewrites the weak lines in place. Point it at a brief and it produces a Requirements Handshake. Point it at research, an…
information-architecture
Map the user journey first, then derive the screen inventory, navigation and hierarchy from it, flagging any screen that serves no journey step. Writes the four-part markdown plus an HTML diagram of the step-to-screen mapping. Use after the brief is settled and before tokens or UI.