Meeting actions sync
Extracts decisions and action items from meeting notes and transcripts, writes them into the tracker with owners and due dates, and chases completion before the next occurrence. Use when the user mentions meeting follow-ups, action items, "who owns this," meeting notes to tasks, or wants recurring meetings to actually produce outcomes.
Meetings don't fail in the room; they fail in the week after, when the decisions evaporate and the same agenda reconvenes. The gap between "we agreed" and "it's tracked with an owner and a date" is where organizational momentum dies. This skill closes the gap the moment the meeting ends — and then does the unglamorous part: it follows up.
When to run
- Post-meeting, on note/transcript availability (recommended auto-trigger for recurring meetings)
- "Pull the action items from the roadmap review" / pre-meeting: "what's still open from last time"
Computer scopes
- Meetings: notes and transcripts for the covered meetings
- Tracker: task creation and status (gated writes)
- Directory: attendee identities for owner resolution
- Chat: the meeting's channel for reminders (gated posts)
Workflow
- Extract from the record: decisions (what was agreed), actions (who does what by when), and open questions. Quote the source line for each — extraction is claim-making, cite it.
- Resolve owners to real directory identities; an action with no ownable owner or no plausible date is flagged back to the organizer, not invented.
- Dedupe against existing tracker items — recurring meetings re-discuss tracked work; attach updates rather than spawning twins.
- Write new items to the tracker (drafted for approval on first runs; L2 with audit once the team promotes it) and post the recap to the meeting channel.
- Chase: reminder to owners at 48h-before-due; unresolved items roll into a "still open from last time" block delivered to the organizer before the next occurrence.
- Decisions (not tasks) get appended to the team's decision log — searchable forever.
Guardrails
- Starts L1 (drafted writes and posts), promotable to L2 per team after four clean runs.
- Chasing tone is neutral and channel-appropriate; two reminders max per item, then it's the organizer's human problem — the skill nudges, it never nags.
- Ambiguity is surfaced, never resolved by invention: no fabricated owners or dates, ever.
Output
Post-meeting: recap (decisions with citations, actions with owners/dates, open questions) + tracker items. Pre-next-meeting: the still-open block. Continuous: the decision log.Eval fixtures
- Set: 12 sandbox meeting transcripts with gold decision/action lists, including 9 ambiguous items (no clear owner or date) and 5 duplicates of tracked work.
- Rubric: extraction recall/precision; all ambiguous items flagged not invented; dedupe correctness; citation validity.
- Ship gate: extraction recall ≥ 0.9; invention rate 0; duplicate-spawn rate 0.
