| name | episode-from-audio |
| description | End-to-end pipeline that takes an edited Bikeshed Podcast audio file (mp3/wav/m4a) and produces (1) a timestamped transcript with speaker labels, (2) a recommended title, (3) a short website description, (4) a tweet-length post, and (5) a full-length show-notes summary. Use whenever Matt uploads or points at a podcast audio file (anything under the-bikeshed-pod or named like an episode) and asks for transcription, episode assets, show notes, a title, a tweet, or "everything for the new episode". Supersedes the VTT-only `episode-summary` skill for the audio-first workflow; wraps it for the asset-generation half. |
Bikeshed Episode From Audio
Matt co-hosts The Bikeshed Podcast (bikeshedpod.com) with Scott and Dillon. After each recording he hands Claude the edited audio file and wants the full publishing bundle back: the transcript itself, plus the four promotional assets the episode-summary skill already knows how to produce.
This skill owns the audio-to-transcript half of the pipeline and then delegates the asset-generation half to episode-summary. Do not re-derive the asset format from scratch — read episode-summary/SKILL.md (in the same .agents/skills/ directory, or via the bikeshed-episode-assets plugin skill) and follow its rules for title / short summary / tweet / detailed summary.
Five deliverables, in this order in the chat response:
- Transcript files written to disk (paths reported in chat)
- Recommended title
- Short summary (the website description)
- Tweet (ends with
bikeshedpod.com, under 280 chars)
- Detailed summary (fenced markdown block)
Step 1: Locate the audio file
The audio is usually in one of:
/mnt/user-data/uploads/ (when uploaded in chat) — common filenames: episode.mp3, bikeshed-NN.m4a, anything ending in .mp3 / .wav / .m4a / .mp4 / .flac / .ogg- The
the-bikeshed-pod workspace folder, often under public/ or a per-episode subdir
If the filename isn't given, list the most likely directory and pick the obvious audio file. If multiple candidates exist, ask which one.
Step 2: Run the transcription script
The script lives at .agents/skills/episode-from-audio/scripts/transcribe.ts. It's a Bun script that uses nodejs-whisper (whisper.cpp under the hood) for transcription and sherpa-onnx-node for pyannote-based speaker diarization, then merges them so each segment has a SPEAKER_XX label.
One-time setup (only if missing)
cd .agents/skills/episode-from-audio/scripts
bun install
brew install ffmpeg
./setup-diarization.sh
nodejs-whisper will auto-download the whisper.cpp model the first time it runs (~1.5 GB for medium). The diarization models are fetched from sherpa-onnx's GitHub releases — no HuggingFace token required, unlike the Python pyannote path.
If the diarization models aren't present, the script still produces a transcript — it just skips speaker labels. Warn Matt in that case so he can decide whether to run the setup script and re-run, or proceed without speakers.
Run it
bun run .agents/skills/episode-from-audio/scripts/transcribe.ts <audio-path> \
--model medium \
--min-speakers 2 --max-speakers 4
Defaults are sane for a typical 3-host Bikeshed episode. Use --model large-v3 for higher accuracy at the cost of time, or --model small for a faster draft. The script normalizes the input to 16 kHz mono WAV via ffmpeg before transcribing/diarizing.
Outputs are written next to the audio:
<stem>.transcript.json — structured segments + speaker list (use this when programmatically reading)<stem>.transcript.vtt — WebVTT with <v SPEAKER_XX> cues; this is what episode-summary expects<stem>.transcript.txt — human-readable grouped-by-speaker transcript
Report the three paths back to Matt before moving on.
Step 3: Map SPEAKER_XX labels to real names
Diarization produces opaque labels (SPEAKER_00, SPEAKER_01, ...). Read enough of the transcript to infer which label belongs to Matt, Scott, and Dillon:
- Matt — works at Anthropic, talks about Claude / Cowork / agent infrastructure, lives in NYC area, talks about his kids and Peloton
- Scott — frontend / React / Next.js leanings, often grounds the conversation in shipping product
- Dillon — infra / runtimes / OSS-governance angles, often the most chronically-online of the three
If two labels look ambiguous, surface your guess and ask Matt to confirm before publishing.
Once mapped, rewrite the .vtt and .txt files in place, replacing SPEAKER_00 etc. with Matt, Scott, Dillon. This makes the files useful as published show transcripts too.
Step 4: Generate the four promotional assets
Switch into episode-summary mode. Read .agents/skills/episode-summary/SKILL.md if you haven't already this session and follow its rules verbatim for:
- Recommended title —
[Memorable hook]: [Plain-English description], hook lifted from a phrase the hosts coined or repeated
- Short summary — 2-4 sentences, opens with "Matt, Scott, and Dillon...", optionally ends with a "Plus..." standup tease
- Tweet — opens with
New episode 🚲, ends with bikeshedpod.com, under 280 chars
- Detailed summary — fenced markdown block,
# title heading, ## sections for main topic / side topics / Standup, bold key terms
When reading the transcript for hook candidates, scan the whole thing — the title-worthy phrase often shows up in the middle or end of the episode.
Step 5: Reply format
Output the assets inline in chat (no file creation for the assets themselves — only the transcript files were written). Use this exact structure so it copy-pastes into the publishing flow:
## Transcript
- JSON: `<path>`
- VTT: `<path>`
- TXT: `<path>`
- Detected speakers: Matt, Scott, Dillon
## Recommended Title
**[Title]**
## Short Summary
[2-4 sentences]
## Tweet
[Tweet copy ending in bikeshedpod.com]
## Detailed Summary
```markdown
# [Title]
[1-2 sentence framing paragraph]
## [Main topic section]
...
## [Side topic section]
...
## Standup / Life Updates
...
```
Voice and editorial notes
Inherit everything from episode-summary: substantive on the tech without being dry, summarize the hosts' takes (don't add your own), a little wit is welcome, bold the terms that scan. If the title or tweet feels borderline, offer one alternative — don't pile on options unprompted.
When to NOT use this skill
- Matt uploads a VTT or text transcript directly (no audio) → use
episode-summary instead, skip the transcription step
- Matt wants only a transcript, no assets → run the script and stop after Step 3
- Matt wants only the assets from an existing transcript → use
episode-summary
