video-research — independently scanned and version-tracked by SaferSkills.
SaferSkills independently audited video-research (Agent Skill) and scored it 100/100 (green). The audit ran 55 deterministic rules across Security, Supply Chain, Maintenance, Transparency, and Community; it found 0 high-severity and 0 lower-severity findings. The full rule-by-rule trace and per-finding evidence are below. Free, methodology-open.
Findings & checks · 0 flagged
Every scanned point with the score it earned and what moved between them.
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
This is a framework template, not drop-in. To use this skill you need to self-build:
~/.claude/wiki/ entities and ~/.claude/projects/{{SCOPE}}/memory/ cluster, the "personal-context analysis" layer will analyze against empty content. You can either build the infra or use this skill purely for NotebookLM CLI baseline summary (skip personal-context section).~/Desktop/projects/{{YOUR_IDEAS_DIR}}/ideas/ exists for output. Adapt to your own ideas directory or override at runtime.This is a framework, not a drop-in tool.
固化的影片研究流程:baseline 摘要由 NotebookLM 自動生成 + Claude 用使用者個人 context 做「對你的意義」分析。
2026-04-26 對「五大 Agent Memory 框架」影片做了三方對比實驗(baseline / NotebookLM v2 strict / subagent independent web research),結論:
詳見 memory/reference_video_research_three_way_pattern.md。
# 從使用者訊息抽 URL(regex 抓 https?:// 開頭)
URL="<paste from user message>"
# 抽影片標題給 notebook 命名
TITLE=$(echo "$URL" | xargs -I{} python3 -c "
import urllib.parse as u
p = u.urlparse('{}')
print(f'{p.netloc} {p.path}'.replace('/', '_')[:60])
")
# 建 notebook
nlm notebook create "Video research: $TITLE $(date +%Y-%m-%d)"
# 從輸出抽 notebook idnlm source add <NB_ID> --url "$URL" --wait --wait-timeout 60如果 type 不是 youtube_video(如 Bilibili 是 web_page),主動告訴使用者「這個平台 NotebookLM 只抓頁面 metadata,不看影片內容;摘要會基於 UP 主寫的描述」。
nlm report create <NB_ID> --format "Briefing Doc" --language zh-Hant --confirm
# 從輸出抽 artifact_id
# poll status 直到 completed(通常 10-30 秒)
nlm studio status <NB_ID>
# 下載
DEST_DIR=~/Desktop/projects/{{YOUR_IDEAS_DIR}}/ideas/$(date +%Y-%m-%d)-video-$(echo "$TITLE" | tr -cd '[:alnum:]-' | head -c 30)
mkdir -p "$DEST_DIR"
nlm download report <NB_ID> --id <ART_ID> -o "$DEST_DIR/briefing.md"讀 briefing.md 之後,主動翻:
~/.claude/wiki/index.md — 找跟主題相關的 wiki entitymemory/MEMORY.md + _index_*.md cluster files — 找相關 user / feedback / project / reference 記憶(兩層 index 結構,2026-05-07 起;MEMORY.md 看到相關 cluster pointer → drill 進對應 _index_<topic>.md 拿 sub-entries → 再 Read 具體 topic file)對話開場給使用者的「對你的意義」分析應該包含(按需要選 3-5 項,不必全寫):
Fact-check 紀律(critical)
寫進「對你的意義」分析的具體技術名詞(framework 名 / API name / feature name / 機制名 / 版本號 / 數據),baseline 提到的當作未驗證 lead,不直接寫進對比表 / 結論。要寫之前必須第一手驗證:
特別針對 Bilibili / 字幕抓不到的平台:baseline 本質是 UP 主描述頁的二次轉述,技術細節錯誤率高(2026-04-26 真實踩坑:把「Letta Sleeptime 異步學習」當 V1 廢的舊機制寫進對比表,第一手查證才發現 sleeptime 跟 heartbeat 是兩個獨立概念,sleeptime V1 後保留並主推;詳見 feedback_no_secondhand_research_as_fact.md)。
不要做的事
只在使用者明確要求時才跑:
Create Your Own + 嚴格 prompt(必須列名 + 等量篇幅 + 禁偷渡,詳見 reference_nlm_deep_research_prompt_dependency.md)reference_video_research_three_way_pattern.md)nlm mindmap/audio/infographic create--cookies-from-browser chrome 都繞不過。Bilibili 走 --url 只抓 UP 主描述,不抓字幕。BibiGPT $19.80/月對中文 platforms 字幕有護城河。nlm research import跑完後告訴使用者:
<path>~/.claude/projects/{{SCOPE}}/memory/reference_video_research_three_way_pattern.md~/.claude/projects/{{SCOPE}}/memory/reference_notebooklm_mcp_cli_repo.md~/.claude/projects/{{SCOPE}}/memory/reference_nlm_deep_research_prompt_dependency.md~/Desktop/projects/{{YOUR_IDEAS_DIR}}/ideas/2026-04-26-video-research-experiment/~30 seconds. Free. No account. Every finding cites a rule and a line of evidence.