plaud-digest

name	plaud-digest
version	1.0.0
description	Summarize multiple Plaud recordings into a digest. Use when the user says 'weekly report', 'digest of this month', 'what meetings did I have this week', 'recap of last quarter', or asks to roll up multiple recordings into one overview.
metadata	{"requires":{"bins":[]}}

plaud-digest

Read plaud-shared first.

When to use

• User asks for a roll-up across multiple recordings.
• Time window is explicit ("this week") or implicit ("recap of recent meetings").
• Scope is "what happened", not "find one specific meeting" (that's plaud-find).

Steps

1. Resolve the window. Use the date interpretation table in plaud-find for relative phrases.
2. List the corpus. list_files with date_from / date_to. Cap at 50 recordings — if the window returns more, ask the user to narrow it.
3. Fetch notes in batch. For each recording, call get_note. Do not call get_transcript unless a specific recording merits a deeper pull.
4. Synthesize. Produce a structured digest:
   • Headline — one-line theme of the window.
   • By recording — one bullet per recording: • name (date, duration) — one-sentence takeaway.
   • Recurring themes — topics that appeared in ≥ 2 recordings.
   • Open action items — aggregated across recordings, deduplicated.
5. Cite sources. Every non-trivial claim must reference the recording it came from, using the file name (not the raw ID unless the user asked).

Budget

• Hard cap: 50 get_note calls per digest. If the window has more recordings, compress or ask user to narrow.
• Skip recordings where note_list is empty — mention them at the end under "unsummarized".

Anti-patterns

• Do not load transcripts just to pad the digest.
• Do not synthesize across windows the user didn't ask for ("while we're at it, here's last month too").
• Do not invent action items that aren't in the notes — only aggregate what's there.