Add "pull YT subs before Whisper" step + footgun

[evnchn]

TL;DR

Add "pull YouTube's own subs before Whisper" to the pipeline (step-2 prelude + one footgun row). yt-dlp --list-subs costs ~5s and pays three ways — most importantly it catches a video that's already in the target language before you download + transcribe it for nothing.

Verdict from a live run: on youtube.com/watch?v=ohCAPyqlJpo (a 16-min English talk), the en-orig tag alone would have short-circuited the whole "localize to English" pipeline. And YouTube's auto-caption vs Whisper large-v3-turbo agreed on 98.2% of words — but the 1.8% delta is exactly where a human is needed, and each engine was right where the other was wrong.

Diff is +2 / −1 lines, matching the skill's dense-footgun house style. Doc-only.

---

The three payoffs (why this earns a line)

1. Language tripwire. --list-subs shows a <lang>-orig track = the video's true source-caption language. If that's already your target (e.g. en-orig), the localize job is a no-op — stop before downloading/whispering. This skill previously had no such tripwire; the trigger video is literally already English.
2. A free, often-better transcript. A manual/creator track in the source language beats Whisper for free: real punctuation, correct proper nouns, no injected [music] tokens.
3. A free second ASR to cross-check. An -orig auto-caption is another ASR — diffing it against Whisper surfaces exactly the lines that need a human fix.

Trap folded into the row: the 100+ other languages in --list-subs (incl. zh-Hant, ja, ko) are YouTube auto-translations, not creator tracks — don't mistake a zh-Hant row for a real Chinese source.

Evidence — YT auto-caption vs Whisper large-v3-turbo (same English audio)

	YouTube auto-cap	Whisper turbo
cues	493	254
words	3006	3010
[music] tokens	58	0
normalized word-seq similarity	—	0.9820

• Segmentation: Whisper's 254 sentence-ish segments are far more dub-ready than YT's 493 rolling display cues (and YT injects 58 mid-sentence [music] tokens). → Whisper is the better base transcript.
• Proper nouns (the 1.8% that matters):
  • Arthur Andersen — YT correct, Whisper wrong (Anderson). Context confirms the Enron-era auditor ("one of the big five", "$25 million in audit fees"), spelled Andersen.
  • Sponsor eSIM brand — both wrong: YT Saylite/Sayle, Whisper Saley. Ground truth is Saily (Nord Security's eSIM, saily.com/invisiblegame, code INVISIBLEGAME) — verified against saily.com.
  • audits ⇄ orders — disagree both directions (context-dependent mishear).
• Remaining delta is ~US/UK spelling (penalized/penalised, color/colour) and number formatting ($95 vs "95 dollars") — cosmetic, not substance.

Takeaway baked into the wording: don't pick one engine blindly — cross-check them; the disagreements are a free proper-noun error detector. Neither engine alone catches "Andersen", and neither gets "Saily".

Repro: yt-dlp --list-subs, pull en-orig, run whisper --model large-v3-turbo --language en, diff normalized word sequences.

🤖 Generated with Claude Code