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

1. Transcribe
2. Run your existing transcript flow for the target video (word-level preferred for reliable speaker calibration).

3. Keep all raw outputs.

4. Speaker cleanup (small-impact)

5. Preserve the default/primary speaker mapping.
6. Only flip clearly short-interjection/question lines when they are obviously misattributed.

7. Keep line count and timestamps unchanged.

8. Translate (optional)

9. Translate content line-by-line.
10. Keep format exactly: [HH:MM:SS] Speaker: text.

11. Keep technical terms unless explicitly asked to simplify.

12. Package for handoff

13. Create concise structure: transcripts/, artifacts/, MANIFEST.txt.
14. Include one reviewed transcript + one translated version when asked.

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.