transcribe

name	transcribe
description	Transcribe a local media file or a YouTube URL into a saved transcript for a Stream Producer project, using OpenAI Whisper directly via a standalone Node.js script (no Java/server dependency). Use when the user asks to "transcribe", "transcribe this recording", "transcribe a YouTube video", or wants a transcript before generating chapters or a summary.
allowed-tools	["Read","Write","Bash","Glob"]

Instructions

Transcribe a local file or YouTube URL for project $ARGUMENTS (project name, followed by either a local file path or a YouTube URL).

Workflow

1. Resolve the project name (see topic-ideas skill for the resolution rule). Create projects/<name>/ if new (the script also creates it, but do this first so you can report the right path even on failure).
2. Determine the input from $ARGUMENTS: a local file path or a YouTube URL. If neither was given, ask the user. Optionally ask for a --language hint (ISO code) if the content isn't in English and they want better accuracy.
3. Confirm prerequisites: .env in the repo root has OPENAI_API_KEY set (point the user at .env.example if not), and ffmpeg/yt-dlp are reachable — the script will fail clearly if they aren't, just relay that message.
4. Run one of:

   node scripts/transcribe.js --project <name> --file <path> [--language <code>]
   node scripts/transcribe.js --project <name> --youtube-url <url> [--language <code>]
   

   This writes directly to projects/<name>/05_transcription_transcript.md — do not read the whole transcript back into context, it can be long. The script only prints a short summary line (segment count, character count, output path).
5. Relay the script's stdout summary to the user. On a non-zero exit, surface the script's stderr message verbatim and stop — do not attempt to fabricate a transcript or retry silently.
6. Note: transcript entries show Speaker: Unknown for every line — plain OpenAI Whisper does not diarize speakers. This is expected, not a bug.