transcribe-podcast

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

1. Transcribe via AssemblyAI (speaker_labels: true, punctuate: true, format_text: true).
   • Save raw output to <source_stem>.raw.md — speaker-labelled prose, no timestamps in this file (timestamps are kept in the sidecar JSON for downstream skills). One paragraph per speaker turn.
   • Also save <source_stem>.assemblyai.json (full API payload) for downstream tools.
2. Produce the reading version at <source_stem>.podcast.md:
   • Remove filler words: um, uh, er, ah, you know (filler use), like (filler use), I mean (filler use), sort of, kind of (filler use), right? (tic).
   • Remove false starts and immediate self-repetitions.
   • Do not paraphrase, summarize, or reorder. Preserve the speaker's meaning, vocabulary, and tone. This is a cleaned transcript, not a rewrite.
   • Preserve speaker labels: **Daniel:**, **Guest:**, etc. If <source_stem>_speakers.json exists, use those names; otherwise use AssemblyAI's A/B/C mapped to Speaker A / Speaker B etc., and ask the user to confirm names if it's a 2+ speaker episode.
   • Add ## section headers at natural topic transitions. Aim for one header every 3–8 minutes of content. Headers should be short, descriptive, sentence-case (## Why diarization matters, not ## Diarization). Do not invent topics — derive headers from what's actually said.
   • Add a single # <Episode Title> at the top if the user supplied one (or it's inferable from the source filename); otherwise skip the H1.
3. Report both file paths to the user, plus a one-line word-count and duration summary.

Output files

File	Contents
<stem>.raw.md	Raw transcript, speaker labels, no editing
<stem>.assemblyai.json	Full AssemblyAI JSON (timestamps, confidences)
<stem>.podcast.md	Reading version — fillers removed, headers added

When to use

• User provides podcast audio and asks for a transcript.
• User says "transcribe this podcast", "transcribe this episode", "give me a podcast transcript".
• User wants both a verifiable raw transcript and a clean reading version (the common case for show-notes / blog repurposing).

When NOT to use

• Single-speaker memo, voice note, or dictation → use transcribe-assemblyai or transcribe-gemini-cleaned directly.
• Meeting / interview where the user only wants timestamps → transcribe-assemblyai.
• User explicitly wants only one output → use the underlying single-output skill.

Notes

• AssemblyAI cost is ~$0.37/hr with speaker labels. Flag if the audio is >4 hours.
• The reading version is derived from the raw transcript, not re-transcribed — one API call total.
• If section-header inference is unreliable on a particular episode (e.g. very meandering conversation), produce the cleaned prose with paragraph breaks only and tell the user headers were skipped.