yt-transcribe

name	yt-transcribe
description	Fetch a single YouTube video's transcript and summarize it. One-shot — paste a URL or video ID, get a journalist-grade summary. Same transcript engine (and WARP proxy support) as the /yt briefing loop; no channels, no state, no rating.
argument-hint	URL or VIDEO_ID. Optional --lang pl|en|auto (defaults to auto).

Step 1 — Fetch the transcript

bun run src/yt-transcript.ts <VIDEO_ID_OR_URL> --lang <lang>

• First argument: URL or VIDEO_ID straight from the user — the script parses both formats (youtube.com/watch?v=…, youtu.be/…, or a bare 11-char ID).
• --lang selects which caption track to fetch (the input), not the summary language. auto by default (best available track); pass pl / en only if the user wants a specific track.
• The transcript goes to stdout; stderr carries diagnostics. Run it bare — no redirects.

Exit codes — the message you show the user must match the real cause. Never report a tooling failure as "no subtitles". When the code is non-zero, read the script's stderr and surface that actual reason; do not paraphrase it away.

• 0 — transcript on stdout → continue to Step 2.
• 1 → the video genuinely has no subtitles. Tell the user exactly that and stop.
• 2 → YouTube rate-limit / IP block (common on datacenter/VPS IPs). Tell the user and stop; it's transient. Recovery: route fetches through a proxy — see docs/warp-proxy.md (set YT_BRIEFING_PROXY).
• 3 → tooling/integration failure (yt-dlp missing, fetch error, unavailable or private video, empty/unparseable track, bad input). This is NOT a missing-captions case. Show the stderr line verbatim so the user sees the real problem, and stop.

---

Step 2 — Summary

Read the transcript from the tool result and respond as a narrative.

• Header: title, author, date
• One introductory sentence
• One paragraph per significant thread — a bold label + continuous prose, concrete facts. Preserve the logic and rhetoric of the original.
• Verdict: what's strong, what's weak, whether it was worth it

Style:

• an intelligent journalist — not like an AI assistant summarizing an article
• no generalities, always specifics: what was the thesis, what was the argument
• if the material has a clear host and guests, show that in the text

Language: Read data/config.json → output_lang once and write the summary in that language — the same language chosen at onboarding that /yt uses. Order of precedence: (1) a language the user explicitly asks for in this request wins; (2) otherwise output_lang from config; (3) only if config is missing, match the language of the video. This is the output language and is independent of --lang (which only picks the caption track to fetch). Don't mix languages: write fully in the target language and insert foreign words only when they are (a) a proper name of a technology/product, or (b) an established technical term with no natural equivalent (API, REST, JSON, webhook, endpoint). Translate everything else.

---

Rules

• Transcripts: never paste the raw transcript into chat — the summary is the artifact.
• One-shot, stateless: this skill reads no channels, writes no state, asks for no rating. It is independent of the /yt briefing loop, though it shares the same transcript engine and proxy. For the recurring channel briefing, use /yt instead.