Watch Video

Watch video — transcribe and analyze any video, at the depth you choose

The general, any-source video tool. Point it at a YouTube link, a Loom, a Vimeo, a Riverside export, a Zoom recording, a webinar, or a competitor's ad — it pulls the transcript, marks the key moments with timestamps, and writes a summary that flags action items, decisions, and quotable lines. v1 is transcript-first; the frame-and-vision roadmap sits under Deferred capabilities.

Relationship to /transcripts

watch-video is the general any-source acquisition-to-summary tool: it fetches a transcript from whatever platform the video lives on, then produces transcript + key moments + summary. /transcripts (transcript-analysis, at primitives/social/youtube/transcripts) stays the YouTube-specific deep insight-extraction pipeline — SCQA structure, the verbatim-quote Iron Law, feeding icp-behavioural + tov-guidelines. They compose, they don't compete: when a YouTube job needs deep structured insight extraction, hand the transcript watch-video pulls to /transcripts. Neither is deprecated.

Triggers

Run when the user says: "watch this video", "transcribe this Loom", "analyze this video", "summarize this recording", "key moments from this", "what happened in this video", "video notes from [url]".

Do NOT run for:

• Deep insight extraction from a YouTube transcript → /transcripts
• Sales-call win/loss analysis → /win-loss
• A 2-3 sentence answer the user could get without artifacts → just answer

Inputs

Required: a video URL (any supported source) or a pasted/linked transcript.

Optional: the video's purpose (client call, competitor ad, webinar, talk) — sharpens the summary framing and the capture routing.

Process (v1 — transcript-first)

1. Parse the source

Detect the source from the URL pattern or file extension: YouTube (youtube.com, youtu.be, /shorts/, raw 11-char id), Loom (loom.com/share|embed), Vimeo, Riverside, a Zoom recording, or a local file. If ambiguous, ask.

2. Pull the transcript

Backend by source, in order:

Clean the pulled text: strip caption tags, de-duplicate rolling captions, paragraph-break on long pauses. Keep the segment timestamps — they carry the whole key-moments layer.

3. Produce the three artifacts

From the timestamped transcript alone — no frames needed:

1. Transcript — cleaned, timestamped, speaker-attributed where the source distinguishes speakers.
2. Key moments — a timestamped list: view or topic changes, the moment a claim lands, anything that reads as a decision, action, or notable event. Each entry cites the transcript.
3. Summary — TL;DR, key moments, action items, decisions, quotes worth keeping, open questions. Full template: the premium reference.

4. Optional — capture to the taste-library

Offer to save the summary for reuse. Routing:

• Competitor / ad / talk / webinar / marketing-reference video → projects/research/taste-library/resources/{MMYY}-video-{slug}.md
• Client call / meeting recording (carries PII) → the client folder, not the shared taste-library, and redact per pii-redaction.md first.

Evidence-bound discipline

This skill is bound by evidence-bound-outputs.md. Every extracted claim — a key moment, an action item, a decision, a flagged quote, a theme — cites a verbatim (or near-verbatim) quote from the transcript with a timestamp and, where the source distinguishes them, the speaker. No quote to back a claim → lower the confidence per ontology.md ([INFERRED] / [ESTIMATED]) or drop the claim. Never invent a quote, a speaker, or a business fact the transcript doesn't state. Clean a garbled auto-caption lightly if you must, flag it (cleaned), but never paraphrase words into someone's mouth.

Citation shape:

"verbatim line from the transcript" — Speaker (if known) [00:12:34]

Deferred capabilities (future)

v1 ships transcript-first. Three capabilities from the source skill are deferred until there's a live use case; the full design is preserved in the premium reference so implementation is a wire-up, not a redesign:

• Local transcription (Whisper). For videos with no platform captions. Deferred — v1 asks for a transcript instead.
• Visual mode (ffmpeg frame extraction + Claude vision). Frames on a per-source cadence, paired with the transcript window, for demos and slide decks where the screen carries meaning the words don't.
• Multimodal mode (Gemini native video / dense vision). Whole-video ingestion for delivery, pacing, and brand / ad-audit reads.

When a capability lands it becomes a selectable depth mode; the user picks depth, and long videos always confirm before any paid frame or vision pass.

Anti-hallucination guardrails

1. Never invent a quote, speaker, name, or number the transcript doesn't contain.
2. Quote verbatim; flag any light cleanup (cleaned).
3. Keep timestamps on every moment and quote.
4. Mark confidence per ontology.md when a claim isn't directly supported.
5. Say "not available" when the transcript is thin — don't fill the gap.

Attribution

Adapts coreyhaines31/makerskills/watch-video (MIT, © 2026 Corey Haines), accessed 2026-07-08. Re-tooled onto our youtube-transcript MCP + Firecrawl; local-transcription/multimodal deferred.