メニュー
This skill should be used when the user asks to "summarize a YouTube video", "summarize this video", pastes a YouTube URL and asks for a summary, says "what is this video about", or asks to "extract a transcript" from a YouTube link.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Run a bounded, fail-closed Claude Code review gate over the current git worktree with a configurable Claude model and strict read-only isolation. Use for fresh-context different-family reviews before commit, including requests for Opus, Fable, Sonnet, or an explicitly selected Claude model; drive fix and re-review rounds until the configured reviewer returns CLEAN.
Use before committing an implementation slice when a configurable different-family reviewer should inspect the slice, or when coordinating up to three fix/re-review cycles with Claude, Codex, or Pi. Claude reviews delegate to the supervised claude-review-loop harness.
Use when you want a fresh-context code-review gate before committing — runs Pi only with the approved OpenAI Codex GPT-5.6 Sol model at high reasoning over the current git diff in a bounded, observable loop, and only proceeds when Pi returns CLEAN. Never falls back to Claude/Anthropic, local models, OpenRouter, or another provider. Drives review → fix → re-review up to a round cap. Triggers: "have pi review this", "pi review before commit", "run the pi review loop".
Use before committing an implementation slice when a configurable different-family reviewer should inspect the slice, or when coordinating up to three fix/re-review cycles with Claude, Codex, or Pi. Claude reviews delegate to the supervised claude-review-loop harness.
Use when the user asks for a compact goal prompt, goal condition, create_goal objective, continuation objective, or second-agent goal statement, especially when the destination has a strict character limit such as 4000 characters. Generate a concise, context-specific goal condition from the current worktree, session context, and relevant planning docs without producing a full implementation or review handoff.
Legacy compatibility alias for claude-review-loop. Use only when an existing workflow explicitly invokes opus-review-loop; delegate the review to the canonical configurable Claude gate with the Opus model.
| name | summarize-youtube |
| description | This skill should be used when the user asks to "summarize a YouTube video", "summarize this video", pastes a YouTube URL and asks for a summary, says "what is this video about", or asks to "extract a transcript" from a YouTube link. |
| allowed-tools | Bash |
Extract transcripts from YouTube videos using the summarize CLI (@steipete/summarize) and produce a concise summary. The transcript is extracted locally and then summarized in-context by Claude.
Use this skill when:
summarize "<URL>" --extract --youtube auto --timestamps to extract the raw transcript.--timestamps output is too long (over ~30k chars), re-run without --timestamps or truncate.# Extract raw transcript (preferred - lets Claude summarize)
summarize "<URL>" --extract --youtube auto --timestamps
# Extract and format as markdown via LLM (alternative)
summarize "<URL>" --extract --format md --markdown-mode llm
# Extract with slide screenshots (requires ffmpeg + yt-dlp)
summarize "<URL>" --slides --extract
| Mode | Behavior |
|---|---|
auto (default) | youtubei -> captionTracks -> yt-dlp -> Apify |
web | youtubei -> captionTracks only |
no-auto | Creator captions only (skip auto-generated) |
yt-dlp | Download audio + local Whisper transcription |
--extract so that Claude performs the summarization rather than an external LLM.--youtube auto unless the user requests a specific mode.--youtube web or --youtube yt-dlp as fallbacks.summarize package; assume summarize is already available on PATH.