Meeting summary

Meeting Summary: turn a recorded Teams meeting into structured notes

You are turning one recorded Teams meeting into durable markdown notes: a meeting note filed under its series folder, plus a per-project doc for each workstream it touched. Everything you write must be grounded in the fetched transcript — the cardinal failure here is fabricating a summary when the transcript is missing or partial, or silently overwriting a note that already exists. If you can't fetch the transcript, say so and stop; if the note exists, ask.

First action on every invocation — the config gate

Before replying or asking anything, resolve the notes root, silently:

1. Read ~/.claude/meeting-summary.local (i.e. $HOME/.claude/meeting-summary.local).
2. A line NOTES_ROOT="<path>" with a real existing directory → proceed.
3. File missing, value empty, or still a <...> placeholder → run Setup, then stop.

Setup: Write this template to ~/.claude/meeting-summary.local (never clobber a real value), tell the user to fill in the path to their clone of the meetings notes repo and save, then stop:

# meeting-summary config — per-user, never committed.
# Set NOTES_ROOT to your local clone of the meetings notes repo.
NOTES_ROOT="<PATH_TO_YOUR_MEETINGS_NOTES_REPO>"

When to use this skill

• User pastes a SharePoint recording link (https://...sharepoint.com/.../Recordings/...)
• User pastes a Teams meeting or join link
• User asks to summarize a meeting by name or date ("summarize this morning's sprint review")
• User wants decisions or action items extracted from a recorded meeting

When NOT to use this skill

• Summarizing an email thread, chat, or document — that's ad-hoc summarization, not this skill; no calendar event or transcript is involved.
• Live/future meetings — there is no transcript yet; say so and stop.
• Editing existing notes by hand ("fix the attendee list in Tuesday's note") — just edit the file directly.

Steps

1. Identify the meeting. Locate the calendar event with outlook_calendar_search:

   • SharePoint recording link: URL-decode the filename; extract the meeting name (text before the -YYYYMMDD_HHMMSS-Meeting Recording.mp4 suffix) and the date. Search by name, narrow by date.
   • Teams join link: search by any keywords in the link, or list events near the implied date and match on the event's onlineMeeting join URL.
   • Name/date: search directly. Multiple plausible matches → present them via AskUserQuestion and stop until answered. Zero matches → report what you searched and stop.

2. Read the calendar event. read_resource with calendar:///events/{eventId}. Record: subject, series name (recurring-series subject; for a one-off, the subject itself), start/end + timezone, attendees, meetingTranscriptUrl, and any agenda in the body.

3. Fetch the transcript. read_resource with the event's meetingTranscriptUrl verbatim (format meeting-transcript:///events/{token}). For a recurring series, append ?start={iso}&end={iso} to scope to the occurrence. Delivered in chunks → fetch all chunks before summarizing. No transcript → do not summarize. Tell the user and offer a skeleton note (metadata only, clearly marked as untranscribed) via AskUserQuestion; only write it if they say yes.

4. Resolve the user's identity. Call get_me once — the returned name/email defines whose action items go under "My Follow-ups". Never assume an identity from the notes or config.

5. Gather prior context. Glob {NOTES_ROOT}/Projects/*.md and the meeting's series folder {NOTES_ROOT}/{Series Name}/*.md; read the docs whose names match projects, products, or prior meetings referenced in the transcript. Use them to separate what's NEW from what was already decided, and to spot prior open items this meeting resolved.

6. Analyze the transcript. Extract: discussion topics (grouped, each with a concise summary), decisions (stated precisely, with who confirmed them), team follow-ups (owner + deadline when stated), the user's own follow-ups (explicit assignments and clear volunteerings), the distinct projects/workstreams discussed, and any callbacks to prior meetings.

7. Write the notes. Read references/note-formats.md for the exact file paths, naming, and markdown templates, then write the meeting note and create/update one project doc per project identified. If the meeting note already exists at the target path, ask (overwrite / skip) via AskUserQuestion before touching it. Project docs are append-only history — add the new entry at the top of Meeting History, check off resolved items, never rewrite prior entries.

8. Report and stop. Use the Output format below. Do not commit or push the notes repo — writing the files is where this skill ends.

Output format

Meeting note: {NOTES_ROOT}\{Series Name}\{YYYY-MM-DD} {Meeting Name}.md
Project docs: {created: X, updated: Y — names}
Summary: {topics} topics, {decisions} decisions, {follow-ups} follow-ups ({n} yours)
Resolved from previous meetings: {list, or "none"}

Rules

What to do

• Ground every claim in the transcript. Decisions verbatim-precise with the decider named; follow-ups only when actually assigned or volunteered.
• Full transcript, always. Never truncate or sample a long transcript; fetch every chunk first.
• Ask at every fork. Ambiguous meeting match, recurring occurrence, existing note at the target path, unclear project grouping → AskUserQuestion, never guess.
• Identity comes from get_me, paths come from the config file. Nothing personal is hardcoded in this skill.

What NOT to do

• NEVER fabricate a summary without a transcript. No transcript means a clearly-marked skeleton (with consent) or nothing.
• NEVER overwrite an existing meeting note unprompted, and never rewrite or delete prior entries in a project doc.
• NEVER commit or push the notes repo — file writes only.
• Don't paraphrase decisions loosely — a softened decision is a wrong decision.

Format discipline

• The deliverable is the note files plus the four-line report. No transcript replay in chat, no narrating each step, no recap of the note contents beyond the counts.