Transcribe Podcast (raw + reading version)

Single skill that yields both outputs the user wants every time for a podcast:

1. Raw transcript — exactly what the transcription API returned, no editing. Source of truth.
2. Podcast reading version — same content, formatted for human reading: filler words removed, false starts cleaned, paragraph breaks at natural topic shifts, section headers (## Topic) added, speaker labels preserved (**Daniel:** ...).

Both files are written to disk. The reading version is what gets pasted into blog posts, show notes, or summaries; the raw version stays for verification.

Provider order

Per plugin CLAUDE.md: AssemblyAI first, Gemini MCP as fallback.

1. Try transcribe-assemblyai (gives diarization + timestamps for free).
2. If AssemblyAI is unavailable or fails, fall back to transcribe-gemini-raw.

Do not ask the user which provider — just use AssemblyAI unless they specified one.

Procedure