菜单
Research topics inside a known YouTube channel without a YouTube Data API key. Use when the user provides a YouTube channel URL/handle and wants to find, rank, shortlist, or investigate videos relevant to a topic using channel metadata and optional transcript text.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
Fetch YouTube transcripts/subtitles into cached local files, including timestamped and timestamp-free text artifacts, without loading the full transcript into the agent context by default. Use when the user provides a YouTube URL or video ID and asks to get, fetch, download, search, excerpt, summarize, or otherwise use a YouTube transcript/captions/subtitles file.
Use when the user asks for ultracode, dynamic workflows, Claude Code /workflows-style orchestration, subagent swarms, parallel agents, multi-agent fan-out, broad migrations, codebase-wide audits, high-risk plans, or complex work that needs independent verification.
Use when stress-testing, speccing, or refining a plan; or when the user asks to be grilled or interviewed in detail about technical implementation, UI/UX, concerns, tradeoffs, scope, rollout, or risks.
Use when running the architect/builder loop — acting as the ARCHITECT and team lead (and creative director for UI work) over any fast builder agent (GPT-5.5 via the Codex CLI, Cursor Composer, Grok, or similar), dispatching the builder directly via its CLI, while the repo's docs/HANDOFF.md holds memory and the human owns the gate calls. Triggers when the user says "architect loop", "be the architect", "drive the loop", "next slice", "rule on the disagreements", "judge the build/results", "write the slice spec", or wants to set up or update docs/HANDOFF.md.
Generate smooth, multi-layer box-shadow CSS that faithfully recreates the classic shadows.brumm.af / Tobias Ahlin layered technique. Use when you need soft, realistic elevation for cards, buttons, modals, overlays, or any UI element instead of a single harsh shadow. Supports the original controls: style (soft/sharp/linear), number of layers, vertical & horizontal distance, blur, base opacity, and shadow color. Produces ready-to-paste CSS or Tailwind arbitrary values.
| name | youtube-channel-search |
| description | Research topics inside a known YouTube channel without a YouTube Data API key. Use when the user provides a YouTube channel URL/handle and wants to find, rank, shortlist, or investigate videos relevant to a topic using channel metadata and optional transcript text. |
Search inside a known YouTube channel with a reproducible local workflow:
yt-dlp.Prefer this over browser automation. Use agent-browser only for manual UI checks or auth/consent fallbacks.
python3 <this-skill-dir>/scripts/youtube-channel-search.py catalog \
'https://www.youtube.com/@CHANNEL/videos' \
--output-dir /tmp/channel-research \
--mode flat \
--limit 300
python3 <this-skill-dir>/scripts/youtube-channel-search.py catalog \
'https://www.youtube.com/@CHANNEL/videos' \
--output-dir /tmp/channel-research \
--mode full \
--limit 140
python3 <this-skill-dir>/scripts/youtube-channel-search.py search \
/tmp/channel-research \
'landing pages storytelling websites' \
--since 2024-06-17 \
--top 20
python3 <this-skill-dir>/scripts/youtube-channel-search.py fetch-transcripts \
/tmp/channel-research \
'landing pages storytelling websites' \
--since 2024-06-17 \
--top 15
python3 <this-skill-dir>/scripts/youtube-channel-search.py rank-transcripts \
/tmp/channel-research \
'landing pages storytelling websites' \
--top 10
catalog --mode flat for a fast title-only index.catalog --mode full --limit N when dates/descriptions matter. Full mode is slower because it visits each video page, but it enables --since and better ranking.search for the metadata shortlist. This prints compact ranked results and stores search-results.json.fetch-transcripts only for likely candidates. It calls the sibling youtube-transcript skill and writes transcript artifacts under transcripts/.rank-transcripts for the final list. It reads transcript.txt, scores local content, and prints snippets.Default output directory:
channel-research/
├── catalog-flat.ndjson
├── catalog-full.ndjson
├── search-results.json
└── transcripts/
└── VIDEO_ID/
├── metadata.compact.json
├── transcript.json
├── transcript.md
└── transcript.txt
yt-dlp is required.YYYY-MM-DD or YYYYMMDD. Records with no date are kept, not discarded, because flat catalogs usually omit upload dates./tmp unless the user asks to persist them in the repo.