メニュー
Use summarize.sh to summarize or transcribe URLs, YouTube/videos, podcasts, articles, transcripts, PDFs, and local files. Use for markdownified YouTube transcripts with summarize --extract --format md --markdown-mode llm.
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
Draft or edit public documentation, product, UI or CLI copy, external drafts, or explicitly requested editorial reviews in clear, plain British English. Do not use for ordinary chat, commentary, status, plans, code-only work, research, internal reviews, worker prompts or evidence gathering.
Use only when the user explicitly invokes $grill-me or asks to be grilled, pressure-tested, challenged, interrogated, questioned first, or aligned before work starts. Broad task-alignment workflow for software, research, writing, planning, purchasing, operations, design, and other generic tasks. Do not invoke proactively; if a task seems underspecified, suggest that the user can invoke this skill without entering the workflow.
Maintain and update fast-moving AI tools in the Nix workspace. Use when asked to refresh, verify, package, install, apply, or compare current versions of Codex, Claude, pi, MCP tools, XcodeBuildMCP, markit/markitdown, qmd, spogo, or other AI developer tools managed through nix-ai-tools, nixos-config, Homebrew tap locks, or ai-stack.
High-quality YouTube transcription from audio using yt-dlp + whisper.cpp through the local summarize CLI. Use when captions are unavailable/untrusted or the user explicitly wants audio transcription. For summarize.sh markdownified YouTube transcripts, use the summarize skill instead.
Write durable Codex session goals. Use when the user asks to write, improve, set, update, or replace a session goal; says the current session goal is too vague, too narrow, too specific, or slop; asks for a long-running/autonomous objective to be framed as a goal; or asks Codex to encode vision, north star, current slice, success criteria, deliverable, CTO briefing, alignment needs, review loop, and completion proof into a session goal. Do not use for ordinary planning, CTO briefs, reviews, or research unless the user asks to create or revise a session goal.
Create or revise RFCs and ADRs using Lawbot-Hub templates, strong in-repo examples, and concrete failure-mode examples. Use when drafting, reviewing, or tightening RFC/ADR documents, or when assessing RFC/ADR quality.
| name | summarize |
| description | Use summarize.sh to summarize or transcribe URLs, YouTube/videos, podcasts, articles, transcripts, PDFs, and local files. Use for markdownified YouTube transcripts with summarize --extract --format md --markdown-mode llm. |
Use the summarize CLI from summarize.sh for URL, file, article, podcast, and YouTube work.
Use this skill when the user says:
summarize "https://example.com"
summarize "/path/to/file.pdf"
summarize "https://www.youtube.com/watch?v=..." --youtube auto
If summarize is not on PATH but Nix is available, use the packaged tool:
nix run github:openclaw/nix-openclaw-tools#summarize -- "https://example.com"
For tweet-style markdownified YouTube transcripts, use the LLM markdown path:
summarize "$YOUTUBE_URL" \
--extract \
--format md \
--markdown-mode llm \
--model google/gemini-3.5-flash \
--plain \
--timeout 10m
Do not add --video-mode transcript for this path. That forces raw transcript extraction and can skip markdownification.
For a raw transcript, use:
summarize "$YOUTUBE_URL" --youtube auto --extract --plain
Add --timestamps when timestamps matter.
Prefer an explicit current model. Do not trust auto if debug output shows a stale model.
Google direct model IDs and Ollama Cloud model IDs are different surfaces; verify the
Ollama Cloud model name when the user asks for Ollama.
--model google/gemini-3.5-flashOPENAI_BASE_URL=https://ollama.com/v1 \
OPENAI_API_KEY="$OLLAMA_API_KEY" \
summarize "$YOUTUBE_URL" \
--extract \
--format md \
--markdown-mode llm \
--model openai/gemini-3-flash-preview:cloud \
--plain \
--timeout 10m
As of 2026-06-02, Ollama Cloud exposes gemini-3-flash-preview:cloud; Google direct exposes gemini-3.5-flash.
Use --debug --metrics detailed when the output is raw or empty. In debug output, confirm:
markdownMode=llmtranscriptRequested=true--video-mode transcript for markdown notes