pod2wiki

name	pod2wiki
description	Scan high-signal podcasts and long-form RSS feeds, summarize them, and write karpathy-claude-wiki compatible source-summary pages.

pod2wiki Skill

Use this skill when the user says "scan podcasts", "刷播客", "刷一下播客", "podcast tracking", "播客追踪", "track podcasts this week", or asks to feed podcasts/blog RSS into a karpathy-style wiki.

How pod2wiki is laid out after install

The one-click installer puts everything under the user's workspace:

<workspace>/
├── tools/pod2wiki/scripts/fetch_podcasts.py
├── config/pod2wiki.config.yaml      # the one the user actually edits
├── config/pod2wiki.env              # contains LLM settings after the user fills the key
├── output/pod2wiki/                 # default local output if no wiki path
└── wiki/sources/                    # if user supplied a wiki, this points there

If the user invokes /pod2wiki, the slash command already has the right paths baked in — just run it.

Steps for the natural-language path (no slash command)

1. Find config/pod2wiki.config.yaml. If missing, the user hasn't installed yet — point them at INSTALL-FOR-AI.md.
2. Run a dry-run first to confirm config parses:

python tools/pod2wiki/scripts/fetch_podcasts.py --config config/pod2wiki.config.yaml --env-file config/pod2wiki.env --days 1 --dry-run

3. Run the real scan with --wiki-out if the user has a wiki:

python tools/pod2wiki/scripts/fetch_podcasts.py \
  --config config/pod2wiki.config.yaml \
  --env-file config/pod2wiki.env \
  --output-dir output/pod2wiki \
  --wiki-out wiki/sources \
  --days 7 \
  --write-insight-log

When an RSS feed only ships a short <description> (Latent Space, many Substack podcasts), fetch_podcasts will automatically download the MP3 enclosure and run faster-whisper (tiny model, first 600s by default) to recover the spoken content. Override with --whisper-model {tiny,base,small,medium,large-v3}, --whisper-clip-seconds N (use 0 for full episode), --whisper-threshold N (auto-transcribe when description is shorter than N chars), or --no-whisper to disable. Audio is cached in output/pod2wiki/transcripts/. If faster-whisper is not installed, the run falls back to the RSS description and emits a [whisper] transcription unavailable warning to stderr.

4. Parse stdout JSON and report to the user:

• items_found
• source_pages_written
• raw_pages_written
• translation_pages_written
• insight_log
• verification_warnings

Reversal Narrative Verification Rule

If verification_warnings is non-empty, do not reuse the flagged bullets in downstream research until the user has checked the raw transcript. These warnings catch LLM-added "X rather than Y" framings that often hallucinate the reversal.