Skill · Workflow
meeting-transcript-to-follow-ups
Pulls decisions, action items, open questions and risks from a pasted transcript, resolves every 'I', 'you' and 'by Friday' to a named person and a real date, and checks each item against a verbatim quote. Drafts the recap email, per-owner messages and tickets for approval; nothing is sent. Use when asked to run the "Meeting transcript to decisions and follow-ups" workflow.
The skill itself. Agents read the name and description first and load the rest only when a task matches.
meeting-transcript-to-follow-ups/SKILL.md
---
name: meeting-transcript-to-follow-ups
description: Pulls decisions, action items, open questions and risks from a pasted transcript, resolves every 'I', 'you' and 'by Friday' to a named person and a real date, and checks each item against a verbatim quote. Drafts the recap email, per-owner messages and tickets for approval; nothing is sent. Use when asked to run the "Meeting transcript to decisions and follow-ups" workflow.
---
# Meeting transcript to decisions and follow-ups
Pulls decisions, action items, open questions and risks from a pasted transcript, resolves every 'I', 'you' and 'by Friday' to a named person and a real date, and checks each item against a verbatim quote. Drafts the recap email, per-owner messages and tickets for approval; nothing is sent.
## Inputs
Ask for any of these that the user hasn't given:
- `transcript` (e.g. "[00:02] Priya: Goal today is the billing page launch date. [00:41] Tom: Migration needs two sprints, so earliest is Jan 12. [01:10] Priya: OK, Jan 12 it is. Tom, can you send the cutover plan by Friday? [01:15] Tom: Yes. [01:30] Lena: Who writes the announcement? [01:34] Priya: Let's park that.")
- `attendees` (e.g. "Priya Shah (Head of Product, chair), Tom Becker (Engineering lead), Lena Ortiz (Marketing), Sam Wu (contract designer)")
- `meeting_goal` (e.g. "Decide the launch date for the new billing page and who owns the data migration")
- `meeting_date` (e.g. "2026-11-16 (Monday)")
- `sender` (e.g. "Priya Shah")
- `tracker` (e.g. "Linear")
## Steps
Run the steps in order. Refer to saved results by name. Don't skip a step that says STOP.
1. **Paste the transcript** — Start: Paste {{transcript}} (or rough notes) with {{attendees}} and roles, the {{meeting_goal}} and {{meeting_date}}. {{sender}} sends the recap; {{tracker}} sets the ticket format. Nothing is sent or created.
2. **Map speakers to attendees** — From {{transcript}} and {{attendees}}, make a table: speaker label as written | attendee | role. Match on names, initials and self-introductions; if unsure, write 'Unknown' and quote the ambiguous line. Then list attendees who never spoke and people mentioned who aren't attendees (they can own tasks, flagged external). If there are no speaker labels at all, say so first; every owner is then unconfirmed. Save the result as `speakers`.
3. **Extract in parallel** — Launch these as parallel subagents in one message, then wait for all of them. Work only from {{transcript}}, using {{speakers}} for who said what. Every item needs a verbatim quote (up to 30 words) with speaker and timestamp if any. Never infer an item nobody said. If none, write 'None found.' Number items with your part's prefix. Save the result as `extracted`.
- **Decisions**: Decisions (D1, D2…): the decision in one sentence, who made or confirmed it, the quote. It needs agreement or a clear call from someone with authority. Ideas, preferences and 'let's think about it' go under 'Not decided' at the end, with quotes.
- **Action items**: Every commitment or assignment (A1, A2…): task as a verb phrase, owner exactly as said ('I', 'you', 'Tom', 'marketing'), due date exactly as said ('by Friday', or 'none'), done criteria from the transcript (or 'not stated'), the quote. Include soft commitments ('I can take a look'), marked soft.
- **Open questions**: Unanswered questions and deferred topics ('let's park this') as Q1, Q2…: the question, who raised it, who was named to answer it if anyone, the quote.
- **Risks and blockers**: Risks, dependencies, blockers and concerns raised (R1, R2…): the risk, who raised it, what it affects if obvious, the quote. Add none of your own.
4. **Resolve owners and dates** — Using {{extracted}}, {{speakers}} and {{attendees}}, for a meeting on {{meeting_date}}:
1. Replace each owner reference with a full name: 'I' = the quote's speaker; 'you' = the person addressed (named in the same or previous sentence, or the previous speaker). Teams stay teams, flagged 'needs a named owner'. Note what each was resolved from.
2. Convert relative dates from {{meeting_date}} to 'YYYY-MM-DD (original phrase)': 'Friday' = the first Friday after; 'end of month' = its last day. 'Next sprint', 'soon', 'ASAP' become 'no date (original phrase)', flagged.
3. Merge duplicates, keeping the clearer quote.
4. End with 'Needs attention': actions with no owner, no date or a non-attendee owner; soft commitments; unassigned open questions; unresolved owners.
One markdown table per section, IDs kept. Save the result as `items`.
5. **Check every item against the transcript** — Check each item in {{items}} against {{transcript}}: is the quote there word for word (allowing for filler words and punctuation), and does it support the item, including owner, date and wording? Mark each Verified, Corrected (fixed to what was said, noting the change) or Removed (quote missing or unsupportive).
Say in one line whether the meeting reached {{meeting_goal}} (reached, partly, not), citing decision IDs.
Return tables of Verified and Corrected items (with owners, dates, done criteria), then 'Needs attention', then 'Removed' with reasons. Save the result as `verified`.
6. **Write the follow-ups** — Launch these as parallel subagents in one message, then wait for all of them. Write from {{verified}} only, keeping item IDs in brackets (e.g. '[A2]') and dates as given. Add no tasks, decisions or dates. Short plain sentences, no pleasantries. Save the result as `followups`.
- **Recap email**: From {{sender}} to all attendees. Subject: 'Recap: <goal in under 8 words> – <meeting date>'. Sections: Outcome (one line on the goal), Decisions, Action items (owner – task – due date), Open questions (and who will answer), Needs attention. Under 300 words.
- **Per-owner messages**: One chat message per owner, under 80 words: their tasks with due dates and done criteria, plus open questions assigned to them. If any action or question has no owner, add a message to {{sender}} listing them and asking them to assign each.
- **Ticket list**: One ticket per action item for {{tracker}}: Title (verb first, under 70 characters), Description (two sentences of context, done criteria, the quote), Assignee, Due date (ISO or blank), Labels ('confirm' for soft commitments). A markdown table, then the same rows as CSV with a header.
7. **Tighten the drafts** — Review loop (max 2 passes): 1. Every action item line has a bracketed ID and an owner name or 'Unassigned'.
2. Every due date starts with YYYY-MM-DD or 'no date'; the original phrase may follow in brackets.
3. Recap email under 300 words; each per-owner message under 80.
4. No filler openers or closers ('Hope you're well', 'Let me know if you have any questions!', 'Great meeting today').
5. The ticket CSV has as many data rows as the ticket table.
6. Nothing under Open questions is described anywhere as decided. Save the result as `followups`.
8. **Approve before sending** — STOP and ask the user to approve before continuing: Review {{followups}} and the 'Needs attention' and 'Removed' lists in {{verified}}. Assign unowned items, then approve, edit or reject. Nothing is sent.
9. **Return the follow-ups** — Finish: Return {{followups}} and {{verified}}.
## Done when
Every step has run, each approval was given, and you've reported what was produced and where it was saved.