YouTube Transcript Pipeline (Lite)
Quick Start
Use this when a user asks for:
- YouTube transcript generation and cleanup
- speaker correction around interview/interjector vs main speaker
- translated transcript output while preserving timestamps/speakers
- ready-to-share folder packaging
Minimal Workflow
-
Transcribe
- Run your existing transcript flow for the target video (word-level preferred for reliable speaker calibration).
- Keep all raw outputs.
-
Speaker cleanup (small-impact)
- Preserve the default/primary speaker mapping.
- Only flip clearly short-interjection/question lines when they are obviously misattributed.
- Keep line count and timestamps unchanged.
-
Translate (optional)
- Translate content line-by-line.
- Keep format exactly:
[HH:MM:SS] Speaker: text. - Keep technical terms unless explicitly asked to simplify.
-
Package for handoff
- Create concise structure:
transcripts/,artifacts/,MANIFEST.txt. - Include one reviewed transcript + one translated version when asked.
- Create concise structure:
What not to do in this skill
- Do not introduce heavy automation assumptions.
- Do not change line/timestamp structure unless user explicitly asks.
- Do not aggressively relabel all speaker turns; prefer conservative flips.
Quality checks
- Validate final line counts against source.
- Confirm active final files are present and clearly named.
- Keep every major handoff stage discoverable in manifest.
Reference
See references/limits.md for quick trade-offs and rollback guidance.