← ClaudeAtlas

watch-videolisted

Actually watch or listen to a video/audio source and hold it in context so you can answer questions about it, summarise it, pull quotes, or act on what it says — a QUICK COMPREHENSION action, not a filing pipeline. Covers the local media hooks Claude cannot otherwise discover: hooks/transcribe.py (long-form audio → text, on-device) and hooks/video-watch.py (the VISUAL channel — slides, code, diagrams, terminal, on-screen text that a transcript throws away). Use when the operator shares a video, recording, talk, screencast, demo, lecture, meeting capture or YouTube URL and wants you to engage with its CONTENT — "watch this", "what does this video say", "does he mention X", "summarise this talk", "is this worth my time", "listen to this recording", "transcribe this", "what's on the slides", "pull the code from this screencast". Also the reference for which tool to reach for: documents → MarkItDown or the document-skills · long audio → transcribe.py · video where the SCREEN carries signal → video-watch.py on top
The-AIOS/aios · ★ 40 · AI & Automation · score 79
Install: claude install-skill The-AIOS/aios
# Watch a video — comprehension, not ingestion Someone shares a video and asks you about it. Two hooks make that possible, and **neither is discoverable on its own** — hooks have no semantic match, so a session asked to "watch this" finds nothing and improvises. That gap is the whole reason this skill exists. ## The boundary — read this first > **This skill WATCHES. It does not FILE.** Its output is *understanding held in the session*, not a vault artifact. No note, no frontmatter, no cross-links, no `_index` update, no action-item routing. Working files go to a scratch path (`/tmp`), never to `vault/`. > > **`/aios:ingest` is the filing pipeline** — it exists precisely so that turning a source into a permanent, wired-up vault note is a deliberate act with conventions. If the operator wants that, **hand off** (*"want me to `/aios:ingest` this so it lands in `reflections/ingests/` with the cross-links?"*) — do not half-implement it here. Two paths, one boundary, no overlap. So the shape of a run is: **get it into context → do what was asked → offer, don't assume.** If the ask was open-ended (*"watch this"* with no question), watch it, then say what it is in two or three lines and ask what they want from it. Don't manufacture a deliverable nobody requested. ## Which tool — the actual decision | The source | Reach for | Why | |---|---|---| | Document (PDF, docx, deck, sheet) | MarkItDown or the `document-skills` | Not this skill's job; those already match semantically | |