← ClaudeAtlas

plaud-to-peoplelisted

Turn Plaud recordings (meetings, networking events, calls) into a clean transcript, who-said-what, episode map per person, and people notes with evidence. Triggers: /plaud-to-people, 'разбери запись Plaud', 'кто был на ивенте', 'process my Plaud recording', 'who did I talk to'.
tonydzi/plaud-to-people · ★ 0 · AI & Automation · score 75
Install: claude install-skill tonydzi/plaud-to-people
# /plaud-to-people: Plaud recording → transcript → speakers → people notes Works with Claude Code, Codex, Cursor, Gemini CLI or any agent that can run shell commands. Built and battle-tested on real networking evenings (5-13 hours of audio a day, 2 languages). ## What you need 1. A Plaud device and account. 2. Plaud's official CLI: `npm install -g @plaud-ai/cli`, then `plaud login`. 3. Optional, for your own transcription and diarization (much better than the built-in one, see Step 2): `pip install faster-whisper soundfile` and a GPU (CPU works, just slower). For speaker diarization: `pip install "nemo_toolkit[asr]"` (Sortformer) or `pyannote.audio`. ## Step 1. Pull the recordings ```bash plaud today # list of today's recordings plaud file <id> # metadata plaud transcript <id> -o rec.txt plaud summary <id> plaud audio <id> # download link, lives ~24h: download right away ``` ⚠️ **`start_at` from Plaud is UTC.** Your photos and calendar are local time. Convert with a named time zone (`America/Los_Angeles`), never a fixed offset: a fixed `-7` is an hour off after DST changes. ⚠️ **One recording ≠ one meeting.** Plaud splits long sessions into files back to back, and one file often holds several different people. ## Step 2. Treat Plaud's transcript as a draft, not as truth Measured on our recordings: the built-in transcript text covered only **7-47% of the recording duration** (no human gold transcript, so not a proven speech