Skill: Transcribe Audio (parent brief)

> Note: In the library pipeline, transcription runs mechanically via ruby lib/buttercut/process_footage.rb transcripts <library> (using TranscribeJob), not by dispatching this sub-agent — so footage analysis comes out identical across models. The WhisperX command lives in exactly one place, lib/buttercut/transcribe_job.rb, which also runs standalone: ruby lib/buttercut/transcribe_job.rb <video_path> <output_dir> <language_code> <whisper_model>. refine_instructions.md remains the playbook for the separate (judgment) refinement step that analyze-video Step 3 dispatches.

Transcribes video audio using WhisperX and produces a clean JSON transcript with word-level timing.

SKILL.md is the parent's dispatch brief. The sub-agent's working prompt lives in agent_prompt.md — inline its contents when launching the Task agent. Don't pass SKILL.md.

Parallelism

Launch at most 2 in parallel. WhisperX is already multithreaded internally (~4 CPU threads via CTranslate2); 2 processes is the throughput-vs-RAM sweet spot on a 16GB Mac.

Inputs to gather and pass inline

The parent reads library.yaml and settings.yaml and passes these values inline in each agent's prompt:

• video_path — absolute path to the video file
• transcript_output_dir — where to write the transcript JSON (e.g. libraries/<library>/transcripts)
• language_code — ISO 639-1 code (e.g. en, es) — parent maps from library.yaml's language name
• whisper_model — model size from settings.yaml (e.g. small, medium, turbo)
• transcript_refinement — boolean from library.yaml. If true, also pass:
• user_context (may be empty string)
• footage_summary (may be empty string)

After the agent returns, update library.yaml with transcript: <filename>.json.

Next step

Once all videos have audio transcripts, dispatch analyze-video for visual descriptions.

Dependencies

WhisperX must be installed. Use the setup skill to verify.