meeting-notes-to-action

Turn raw meeting notes or transcripts into structured action: decisions log, action items with owner and deadline, open questions, and a distributable summary email. Extracts who-committed-to-what with date parsing (natural language dates resolved against the meeting date), deduplicates, tracks carryover between meetings, and generates per-owner task lists. Use when the user has meeting notes/transcripts and needs minutes, action items, or a follow-up summary.

Meeting Notes → Action 📋

Convert messy meeting notes into structured, accountable output: decisions made, action items with owner + deadline + confidence, open questions, and a ready-to-send summary email. Includes a deterministic extraction engine for clearly-stated commitments ("Sarah will send the report by Friday") plus carryover tracking between consecutive meetings.

Overview

The universal post-meeting ritual: someone scrolls back through 40 minutes of chat-log-style notes trying to figure out what was actually decided and who promised what. Action items live in five people's heads and evaporate. This skill systematizes it:

• meeting_extractor.py — deterministic NLP over raw notes:
  • Action items: subject-verb-commitment patterns ("X will do Y by DATE", "assign to X", "action item:", checkboxes, "follow up")
  • Owner resolution: names + @mentions + role aliases ("PM", "Sarah")
  • Date parsing: natural language ("by Friday", "next Tuesday", "end of month", "in two weeks") resolved against the meeting date
  • Decisions: "decided/agreed/going with", "approved", "rejected"
  • Open questions: "question:", "?", "TBD", "discuss next time"
  • Dedup: similar items merged; question-vs-action discrimination
  • Carryover: compare with a previous meeting's JSON — unfinished items roll forward with age and staleness warnings
• Outputs: terminal digest, JSON, Markdown minutes, and a summary email draft with per-owner action lists

When to Use

• "Here are my meeting notes — pull out the action items"
• Weekly team meeting → minutes + assignment emails
• "What did we decide last week?" — decisions log from stored JSON
• Tracking recurring meetings: carryover of unfinished items with aging
• Post-standup: converting chat-style notes into tickets

Don't use for: verbatim transcription (needs a transcription tool first), calendar scheduling, or project management sync (Jira/Linear have their own skills) — this produces the structured intermediate those tools consume.

Quick Start

bash

# Basic extraction
python3 scripts/meeting_extractor.py notes.txt --meeting-date 2026-08-12

# Full outputs: JSON + Markdown minutes + email draft
python3 scripts/meeting_extractor.py notes.txt --meeting-date 2026-08-12 \
  --title "Product sync" --json out.json --minutes minutes.md --email email.md

# Weekly carryover: items from last week's JSON that didn't complete
python3 scripts/meeting_extractor.py notes.txt --meeting-date 2026-08-19 \
  --previous lastweek.json --minutes minutes.md

Notes formats accepted: free text, chat-style lines ("10:42 Sarah: I'll send the deck"), checkbox lists, "Action items:" sections — commonly mixed.

Extraction Patterns

Signal	Example	Extracts
<name> will <verb>	"Sarah will send the report"	action, owner Sarah
@name mention + commitment	"@tom can you review?"	action, owner tom (question form)
by <date> / due <date>	"by Friday", "due 2026-08-20"	deadline
action item[s]: section	"Action items: - ..."	everything listed
checkbox	"- [ ] call vendor"	action, unassigned
assign(ed)? to <name>	"assigned to Priya"	owner
decided|agreed|going with	"we agreed to postpone launch"	decision
TBD|question|?	"pricing TBD"	open question

Ambiguity handling: "we should probably" → low-confidence (flagged for review); imperatives without owner → owner Unassigned.

Common Pitfalls

1. Trusting extraction as ground truth. The engine is deterministic, not telepathic — review the confidence flags before sending the email out.
2. Relative dates without --meeting-date. "By Friday" is meaningless without knowing which Friday; pass the meeting date or dates resolve to today's context (and get flagged).
3. Chat transcripts with overlapping speakers. Timestamp lines are stripped; speaker prefixes (Name:) become owner candidates automatically.
4. Sending the full minutes to everyone. The email draft contains per-owner sections — trim for exec recipients; keep decisions prominent.
5. Ignoring carryover staleness. An item carried 3+ meetings is either blocked (needs escalation) or dead (kill it). The report flags both.
6. Mixing decisions with discussion. "We discussed X" is not "we decided X" — only commitment verbs produce decisions.

Verification Checklist

• Every action item has: description, owner (or Unassigned flag), due date where stated
• Relative dates resolved against the correct meeting date
• Low-confidence items reviewed by a human before distribution
• Carryover (if any) checked — completed items marked, stale items escalated
• Email draft's per-owner sections match actual attendees

References

• references/extraction-patterns.md — full grammar, confidence scoring, edge cases
• references/minutes-templates.md — minutes + email formats for different meeting types