Meetings (Beta)

View saved version →

---
name: one-on-one-prep
description: Prepare for upcoming one-on-ones using the user's connected Calendar and recent one-on-one records in Meetings. Use for 1:1 prep with follow-ups, suggested agenda items, and sources.
---

# One-on-one prep

This workflow is read-only. If Calendar or Meetings is unavailable, explain which part of the prep cannot be determined and stop.

## Find upcoming one-on-ones

Use the user's requested day. Otherwise, in the Calendar timezone (or thread timezone if unavailable), use today before 6:00 PM on weekdays, the next weekday at or after 6:00 PM, or the next Monday on weekends. For today, search from now to day-end unless the user requests all of today; for a future day, search the full day.

Discover an event-list/search tool from the user's actually connected Calendar connector and follow its schema and pagination. Do not assume a provider or connector ID, or substitute Meetings' upcoming plan. If multiple calendars are plausible and context does not identify one, ask the user to choose. Limit Calendar reads to the selected day's events and their details; do not look back for history.

Candidates must contain the current user and exactly one other attendee. If attendees are missing from the list, fetch likely events' details before classifying them. Use recurrence metadata when available as context, not as evidence of prior one-on-one notes.

## Find relevant history

Use [$Meetings](app://connector_openai_chatgpt_meetings) `find_all` for the last month, with the other attendee's name or email in `attendee_names` / `attendee_emails` and dates in `start_time` / `end_time`, following the available schema. Titles, topics, and keywords are model-side selection criteria, never `find_all` parameters or attendee values. Follow returned `cursor` pagination within that month as needed to identify the newest relevant records, preserving the default page size and maximum limit.

Confirm prior one-on-ones using titles, participants, and likely candidates' notes when needed. Exclude group meetings and unrelated history; ask the user to choose if ambiguity remains. Fetch selected records with `note_id=<find_all.id>`, normally the latest 2 relevant meetings per person, reusing notes already fetched. A specific meeting identifier supplied by the user can bypass discovery.

## Prepare the brief

For each person with at least one fetched prior one-on-one record, give the upcoming time, up to 3 follow-ups, up to 3 suggested agenda items, and at most 2 source meetings, newest first. Use fetched notes/transcripts as factual grounding and distinguish suggestions from recorded commitments. If the sources support no specific follow-ups or agenda, say so. Keep transcript quotes to short excerpts needed to resolve ambiguity.

Link source titles using this format:

```markdown
[<meeting title>](codex://mcp-app/chatgpt-meetings@openai-curated-remote/meetings.open/meeting/<note-id>)
```

Use the exact durable private `id` returned by `find_all` or `fetch`, unchanged, rather than an HTTPS permalink or provider join URL. Preserve HTTPS permalinks for browser or external-system workflows.

Briefly list skipped people and upcoming times when no relevant prior one-on-one record could be fetched; do not generate prep for them. If nobody qualifies, say that no upcoming one-on-ones had relevant prior notes.

Referenced files: 1

View saved version →

---
name: review-meeting-follow-ups
description: Review a selected day of Meetings records for source-backed commitments, owners, deadlines, decisions requiring follow-up, and unresolved questions. Use when a user asks for an end-of-day meeting review, daily action-item digest, commitments from today's meetings, outstanding follow-ups, or a scheduled weekday meeting recap.
---

# Review meeting follow-ups

Produce a read-only daily follow-up digest from fetched Meetings records. Do not create tasks, send messages, change meeting settings, or write back to any system.

## Workflow

1. Resolve the review day.
   Use the date the user names. Otherwise use the current calendar day in the user's timezone when available, falling back to the thread timezone. Convert the day's start and end to timezone-aware ISO timestamps before searching.

2. Find that day's meeting records.
   Use [$Meetings](app://connector_openai_chatgpt_meetings) `find_all` with the day bounds. Follow pagination when needed. Do not substitute Calendar events for meeting records.

3. Fetch the source material.
   Fetch each result whose `fetch_status` is `fetchable` with `note_id=<id>`, newest first, up to 25 meetings. If more results exist, state how many were not reviewed. Record non-fetchable results and fetch failures under `Processing gaps`; do not invent content for them.

4. Extract only supported follow-up signals.
   Use fetched notes and transcript entries to identify:

   - explicit commitments or assigned action items;
   - an explicit owner and deadline, when stated;
   - decisions that require a concrete next step;
   - unresolved questions or blockers someone agreed to revisit.

   Treat a suggestion, topic, or aspiration as a commitment only when the source clearly assigns or accepts follow-up. Use transcript evidence when speaker attribution or exact wording matters. If a transcript is preliminary, label any transcript-only item as preliminary.

5. Attribute and deduplicate conservatively.
   Resolve the current user only from authenticated session context or an explicit user statement, never from attendance or speaker order. Put an item under `Your follow-ups` only when its owner unambiguously matches that identity. If current-user identity is unavailable, omit `Your follow-ups` and keep the stated owner under `Other owners / waiting on`, or use `Owner unclear`; never guess from an ambiguous pronoun. Merge repeated statements of the same item across meetings, but preserve every supporting source. Treat a later status as superseding an earlier one only when the later source explicitly updates it.

6. Return the digest.
   Keep every item traceable to a fetched meeting title and date or time. Preserve stated deadlines exactly. Write `No deadline stated` when none appears. Mark synthesis beyond an explicit source statement as `Inference`.
   Render each available source meeting title with the following Markdown template, using its exact durable private note ID from the `id` returned by Meetings unchanged.

   ```markdown
   [<meeting title>](codex://mcp-app/chatgpt-meetings@openai-curated-remote/meetings.open/meeting/<note-id>)
   ```

## Output

```text
# Meeting follow-up review for <date>

## Your follow-ups
- <commitment> — <deadline or No deadline stated>
  Source: <meeting title>, <time>

## Other owners / waiting on
- <owner or Owner unclear>: <commitment> — <deadline or No deadline stated>
  Source: <meeting title>, <time>

## Decisions needing follow-up
- <decision and concrete next step>
  Source: <meeting title>, <time>

## Unresolved questions
- <question or blocker and stated owner, if any>
  Source: <meeting title>, <time>

## Processing gaps
- <meeting title>: <not fetchable, pending, incomplete, or failed reason>

## Meetings reviewed
- <time>: <meeting title>
```

Omit empty action sections. Omit `Processing gaps` when there are none. If no meeting records exist, say so and stop. If records exist but contain no supported follow-ups, say that no explicit commitments or unresolved follow-ups were found and list the meetings reviewed.

## Rules

- Always use `find_all` before `fetch` unless the user supplies an exact note or Meeting identifier.
- Stay within the selected calendar day; do not pull older meetings merely to fill an empty digest.
- Prefer calendar meeting time in returned meeting metadata, then recording time as a fallback.
- Do not claim an owner, deadline, completion state, or decision that the fetched source does not support.
- Do not quote more transcript than needed to disambiguate an item.
- Do not execute follow-ups or create another automation. This skill only reports what the source meetings support.
- Use Meetings app deep links for meeting titles in Codex; leave HTTPS permalinks unchanged for any workflow that explicitly requires a browser or external-system URL.

Referenced files: 1

View saved version →

---
name: search-meetings
description: Find recent meetings by date, participant, title, or topic and answer questions about their discussions, decisions, and action items from meeting notes and transcripts.
---

# Search meetings

## Find the relevant records

Use [$Meetings](app://connector_openai_chatgpt_meetings) for `find_all` and `fetch`. Start with `find_all` unless the user supplied an unambiguous meeting or note identifier.

- Use the requested date range and participants; `recent` defaults to the last 7 days. Pass dates as `start_time` / `end_time` and participants as `attendee_names` / `attendee_emails`, following the available tool schema. Titles, topics, and keywords are model-side selection criteria, never `find_all` parameters or attendee values.
- Follow returned `cursor` pagination within that scope as needed before concluding there is no match. Preserve the tool's default page size and maximum limit; do not silently widen the date range.
- Select candidates using titles, times, and attendees. Topic matches need not repeat the user's wording. If metadata is insufficient, fetch only likely candidates' notes to verify relevance before reading transcripts deeply. If ambiguity remains, show up to 5 candidates with title, time, and match reason, and ask the user to choose.
- Fetch selected records using `note_id=<id>` from the results, reusing any content already fetched. Prefer notes for summaries, decisions, and action items; use transcripts for speaker-specific questions, quotes, and detailed chronology.

## Answer with sources

Answer the user's question directly from fetched content, distinguishing inference from recorded facts. State any missing content or conflicts between notes and transcript; prefer the transcript for exact wording. Keep quotes short and relevant, and keep separate meetings' timelines clear.

Use calendar-scheduled meeting times when available, falling back to recording or transcript times. Link every source meeting title using this format:

```markdown
[<meeting title>](codex://mcp-app/chatgpt-meetings@openai-curated-remote/meetings.open/meeting/<note-id>)
```

Use the exact durable private `id` returned by `find_all` or `fetch`, unchanged. Do not substitute an HTTPS `permalink` or derive an ID from a provider join URL. Preserve HTTPS permalinks unchanged for workflows that need browser or external-system links.

Referenced files: 1