Imported from JYeswak/grok_bot_playground (plugin/skills/transcript-extract/SKILL.md). Install upstream with npx skills add JYeswak/grok_bot_playground --skill transcript-extract. Copyright stays with the author.

Transcript extract

Turn a video URL into a transcript plus a short brief, with the chain from URL to transcript to brief unbroken and inspectable. This skill exists because video briefing is a repeat community pattern in the corpus (a small cluster of Bots build on transcript fetching with no shared plugin) and because a brief whose quotes cannot be traced back to a timestamped line is indistinguishable from invention.

The rule that matters more than the brief

A brief without a chain is a rumor. Every claim in the brief traces to timestamped transcript lines, every transcript traces to its source URL plus retrieval method plus retrieval date, and anything unretrievable, inaudible, or auto-translated is labeled as such — never filled in from memory, from a related video, or from listicle knowledge about the topic. Community transcript patterns are community-claim; only the retrieved lines are evidence.

When to use

• Asked to summarize, brief, or quote a video given its URL.
• Asked to brief two videos (compare, contrast, or brief each separately) with each video's chain kept distinct.
• Asked what a video actually says versus what its title or comments claim.
• Do not use for audio-only uploads you own, for transcribing a private meeting, or for editing or publishing video — this skill reads public captions and drafts briefs only.

Inputs and access

• Video URL(s) — one URL for a single brief, two URLs for a side-by-side. Normalize each first: canonical video id, watch URL as given, uploader and title as displayed on the day of retrieval.
• Caption track choice — prefer human-uploaded captions for the requested language; fall back to auto-generated only when labeled as auto-generated in the chain. Never silently mix two tracks.
• Access — public caption endpoints only, with whatever the Bot can already read. If a video needs a login, a rental, or a credential you were not given, record it as unreachable and move on. Do not improvise access.

Sequence — what to do, in this order

1. Normalize the URL — canonical id, full URL as given, uploader, title, duration, retrieval date. One source block per video; with two videos, keep the blocks and everything downstream separate.
2. Fetch the transcript — human captions first, auto-generated only as a labeled fallback. Record method (which track, which language, auto or human) and timestamp range covered.
3. Validate the lines — check the validation section below before writing a single brief sentence. A transcript that fails validation is a finding ("no captions", "track is machine-translated", "only first half covered"), not a quiet gap to write around.
4. Draft the brief — short: what the video argues or shows, in the owner's words where it matters, with every load-bearing claim carrying a timestamp citation. Quote verbatim; paraphrase only with the cited lines alongside.
5. Mark the gaps — inaudible spans, uncaptioned music or demo segments, and anything the brief deliberately omits. End with the single most consequential point per video and what it would take to verify it further.

Validation

• Transcript present for the claimed span; missing captions fail, never silently narrow, the brief.
• Quotes verbatim against the retrieved lines — no cleaned-up wording passed off as a quote, no timestamps invented or rounded beyond the line they came from.
• Track identity stated: human or auto-generated, which language, and whether machine-translated. An auto track never briefs with the confidence of a human one.
• Two-video briefs keep chains separate: no quote from video A cited under video B, no shared claim without a citation to each side.
• Anything unheard, uncaptioned, or unretrieved is labeled UNKNOWN with the reason, never smoothed over.

Output

Per video: a source block (canonical id, URL as given, uploader, title, duration, track identity, retrieval date), the timestamped transcript excerpt that carries the brief, the brief itself with inline timestamp citations, and a gaps list. With two videos, deliver two chains plus a short comparison that cites each side separately. End with the single most consequential point and the decision or follow-up it needs — not a wall of quotes.

Boundaries

• Read-only plus drafts. Never download the video, upload or publish anything, post a comment, or spend on transcription or enrichment while briefing.
• Approval boundary: this skill drafts briefs for the owner to use. Publishing a brief, quoting it publicly, or acting on a consequential claim belongs to a routine with its own approval — nothing here sends or publishes.
• Never present community transcript lore (which endpoints work, which videos "always have captions") as vendor truth. State what worked today, for this URL, with this method.
• Never paste credentials, tokens, or account identifiers into a transcript, brief, or finding. If a source needs access you lack, record it as unreachable and move on.