meeting-notes
/meeting-notes
Zamienia spotkania z kalendarza i transkrypcji notetakerów w notatki Markdown, sortowane według projektów, z notes.md i transcript.md.
- Autor
- Andrew-eGov ↗
- Repozytorium
- meeting-notes-agent ↗
- Licencja
- MIT
- Ostatnia zmiana repozytorium
- 29.10.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 10
- Skrypty
- 0
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 0
Brak sygnałów. Repozytorium zmienione 29.10.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
Co robi i kiedy
Zamienia spotkania z kalendarza i transkrypcji notetakerów w notatki Markdown, sortowane według projektów, z notes.md i transcript.md.
Oryginalny opis (EN)
Turn meetings into structured Markdown notes filed by project. Finds meetings from any connected calendar (Google Calendar, Outlook/Microsoft 365) or straight from transcripts, pulls transcripts from any connected notetaker (Granola, Fireflies, Otter, Fathom, Zoom, ...) or local open-source notetaker (Muesli), from documents attached to the calendar invite, or from transcript files dropped in an inbox folder, then writes notes.md + transcript.md into Meetings/<folder>/ and keeps index.md and each _project.md current. Works with no connectors at all (inbox files only). Use for triggered or sche
Jak działa
- Supporting files
- Workspace
- Modes
- Mode: Init
- Preflight (every mode)
- Mode: Quick
- Mode: Run
- Mode: Setup
- Mode: Review
- Mode: Status
Podgląd SKILL.md
---
name: meeting-notes
description: Turn meetings into structured Markdown notes filed by project. Finds meetings from any connected calendar (Google Calendar, Outlook/Microsoft 365) or straight from transcripts, pulls transcripts from any connected notetaker (Granola, Fireflies, Otter, Fathom, Zoom, ...) or local open-source notetaker (Muesli), from documents attached to the calendar invite, or from transcript files dropped in an inbox folder, then writes notes.md + transcript.md into Meetings/<folder>/ and keeps index.md and each _project.md current. Works with no connectors at all (inbox files only). Use for triggered or scheduled runs, when the user types /meeting-notes (optionally with quick, reconcile, a date, range, init, setup, review or status), or asks to process, file or catch up on meeting notes.
---
# Meeting Notes Agent
One job: every meeting the user attended becomes a structured note, filed in the right folder. Nothing else.
This skill is tool-agnostic. It never assumes a specific connector, operating system, folder or account. It discovers what is available, uses the best option, and degrades gracefully.
Paths in this file:
- `<skill>/…` = this skill's own directory (where this SKILL.md lives).
- `<ws>/…` = the user's **workspace**, the folder holding their meetings (resolved below).
## Supporting files
| File | Read when |
|---|---|
| `<skill>/references/storage.md` | When `<ws>` is on a cloud drive (`drive:…`); every file operation then follows it |
| `<skill>/references/providers.md` | Preflight: finding the calendar and transcript sources |
| `<skill>/references/note-template.md` | Writing any `notes.md` or `transcript.md` |
| `<skill>/references/routing.md` | Choosing a folder; Setup and Review modes |
| `<skill>/references/indexes.md` | Updating the ledger, `index.md`, monthly logs, `_project.md` |
| `<skill>/templates/*` | Creating a new workspace (Init) |
## Workspace
Resolve `<ws>` in this order and use the first that applies:
1. A path the user gave in this request (e.g. `/meeting-notes --dir ~/Notes/Meetings`).
2. The environment variable `MEETING_NOTES_DIR` (check with `echo "$MEETING_NOTES_DIR"`).
3. `Meetings/` in the current working directory.
`<ws>` can be a local path (including a folder synced by Google Drive, Dropbox or OneDrive) or `drive:<folder id or path>` for a Google Drive folder reached through the Drive connector. For `drive:`, read `<skill>/references/storage.md` first and use its operations for every read, write, move and delete in this skill: "delete" means trash, and "rewrite a file" means create the new version, then trash the old one.
If `<ws>/config.md` doesn't exist, run **Init** first. If it exists but other parts of the layout are missing (`routing.md`, `_pending.md`, `index.md`, `_unsorted/`, `_log/`, `.ledger.jsonl`), create each one from `<skill>/templates/` when it's first needed. Never overwrite existing files.
**No writable local filesystem** (e.g. claude.ai or a cloud routine): if a Google Drive connector is available, use a `drive:` workspace (default `drive:My Drive/Meetings`) and everything works as normal. Otherwise you can still process meetings, but you can't file them persistently: produce each `notes.md` as a downloadable file, skip the indexes and the ledger, and tell the user once that persistent filing needs a local folder or a connected Drive.
### Workspace layout
```
<ws>/
├── config.md settings (user-editable)
├── routing.md folder rules (user-editable)
├── _pending.md items waiting for the user's review
├── index.md open action items + last 30 days (generated)
├── _log/YYYY-MM.md full monthly meeting lists (generated)
├── .ledger.jsonl one line per processed meeting (machine state)
├── .schedule.json upcoming post-meeting checks, read by the trigger script
├── _inbox/ transcript files dropped by the user or other tools
├── _unsorted/ meetings with no clear folder
└── <Folder>/
├── _project.md folder overview (generated)
└── YYYY-MM-DD_slug/
├── notes.md
└── transcript.md
```
## Modes
| Invocation | Mode |
|---|---|
| `quick` (the trigger script, after a meeting ends or a file arrives) | **Quick**: just-ended meetings and new transcripts only |
| `reconcile`, no argument, or a daily schedule | **Run** (reconcile): today + catch-up window |
| A date (`2026-10-27`) or range (`2026-10-20..2026-10-24`) | **Run** for those days |
| `init` | **Init**: create the workspace |
| `setup` | **Setup**: first routing proposal |
| `review` | **Review**: pending folders and moves |
| `status` | **Status**: report sources, last run, next checks and pending items; change nothing |
**Unattended** = a scheduled or headless run (no user in the conversation). It must never stop to ask a question; it queues anything that needs the user in `<ws>/_pending.md`. When a run is **on demand** and `_pending.md` has open items, offer a Review first.
**One run at a time.** Quick and Run use the lock `<ws>/.run.lock`:
- Started with `--locked`: the trigger script already holds the lock for you. Don't check it, create it or remove it.
- Otherwise: if the lock exists and is less than 30 minutes old, say a run is already in progress and stop. If not, create it (replacing a stale one), and remove it when you finish, even if the run failed.
Status, Setup and Review don't use the lock.
## Mode: Init
1. Create `<ws>` and copy everything from `<skill>/templates/` into it (`config.md`, `routing.md`, `_pending.md`, `index.md`). Create the empty folders `_inbox/`, `_unsorted/`, `_log/` and an empty `.ledger.jsonl`.
2. Fill in `owner_name`, `owner_email` and `timezone` in `config.md` if a calendar is available. Otherwise ask the user (on demand) or leave them blank (unattended).
3. Tell the user where the workspace is and what sources were found (see Preflight). Suggest `setup` once a calendar is connected.
Init never overwrites existing files.
## Preflight (every mode)
0. **Clock.** Work out "now" before anything else: use the `--now <ISO time>` argument if there is one, otherwise run `echo "${MEETING_NOTES_NOW:-}"` and use that if it isn't empty, otherwise use the system clock (`date -u`). A set time (used by tests and replays) counts as the current time everywhere: "today", which meetings have ended, the checks in `.schedule.json`, and every timestamp you write.
1. **Resolve `<ws>`** and read `config.md`. Fall back to the defaults in `<skill>/templates/config.md` for any missing key, so older workspaces keep working after updates.
2. **Discover sources** with `<skill>/references/providers.md`:
- a **calendar** source (optional),
- **transcript sources** in the order listed under `sources` in config (by default: any connected notetaker, then local notetaker apps such as Muesli, then documents attached to the calendar invite, then the `_inbox/` folder).
3. **Pick the meeting list strategy:**
- Calendar found → **calendar-first** (the calendar decides which meetings happened).
- No calendar → **transcript-first**: every transcript found for the day is a meeting. Title, time and attendees come from the transcript source's metadata or the file. There are no placeholder notes in this mode, since nothing tells you about unrecorded meetings.
- Nothing at all (no calendar, no notetaker, empty inbox) → report that nothing was found, say how to add a source, and stop.
4. Read `routing.md` and the ledger.
Never fail the whole run because one source errors. Note the error, continue with the other sources, and include it in the summary.
## Mode: Quick
Fast, frequent and event-driven. The trigger script starts it only when there's something to do, so it should stay short.
Candidates are only:
- calendar meetings from today that have ended and don't have a complete note yet (this covers every check that came due in `.schedule.json`);
- meetings that a **new transcript** belongs to: inbox files, recordings the notetaker added today, and new attachments on today's invites. In transcript-first mode, each new transcript is a meeting;
- today's `placeholder` and `provisional` notes, which are re-checked for a verbatim transcript.
Then apply the same filters and the same per-meeting steps as Run (below), followed by the indexes and the schedule. Skip the catch-up window and leftover-inbox reporting; reconcile covers those. If nothing was found, change nothing except `.schedule.json`, send no notification, and end with one line ("Nothing new").
## Mode: Run
Work out the days in scope: today plus the previous `catch_up_days` (or the requested date/range). Use `timezone` from config, or the calendar's timezone, for "today" and all dates.
Build the list of candidate meetings for those days (calendar-first or transcript-first). Drop the following:
- events the owner declined or didn't respond to, all-day events, out-of-office, cancelled, events with no other attendees, and events that haven't ended yet;
- **sensitive meetings**: the title or description contains any `sensitive_keywords` entry (case-insensitive, whole word). Write nothing for them, not even an index entry. Count them as "skipped (private)". Any transcript that would match a sensitive meeting (a connector recording or an `_inbox/` file) is **reserved**: never pair it with another meeting, never copy it, and leave inbox files where they are. In the summary, report only the count ("1 private transcript left in the inbox"), never the file name or title.
- meetings already in the ledger with `status: complete`. (`placeholder` and `provisional` meetings stay candidates until the catch-up window passes.)
Process the oldest first, up to `max_meetings_per_run`. If more remain, leave them for the next run and say so in the summary; the catch-up window will pick them up.
For **each** meeting, finish all of these steps before starting the next one, so an interrupted run can resume safely:
1. **Meeting id.** It must be unique **per occurrence**:
- calendar: the event's instance id. If the provider only gives a series-level id (e.g. an ICS `UID` with an `RRULE`), use `<id>_<start as YYYYMMDDTHHMMSSZ>`, and put the series-level id in `series_id`;
- transcript-first: `<source>:<source id>`, or `file:<filename>` for inbox files.
2. **Transcript.** Try each transcript source in config order and stop at the first match. A transcript matches when its start time falls within the meeting's **start − `match_window_minutes`** and **end + `match_window_minutes`**. If several match, take the one that starts closest to the meeting's start, then the closest title. See `<skill>/references/providers.md` for each source type, including how to name the `source`.
No match → placeholder note (calendar-first only).
If the only match is a **fallback**, meaning dictated notes (`muesli:dictation`, an inbox dictation) or a notes document without a verbatim transcript, and a verbatim transcript could still arrive (e.g. the notetaker's recording exists but is empty, or the meeting is still inside its post-meeting checks), write the note with `status: provisional`.
3. **Folder.** If the ledger already has this meeting as a placeholder, keep its folder. Otherwise pick one with `<skill>/references/routing.md`. If that creates a folder, write its rule to `routing.md` and its line to `_pending.md` **now**, before writing the meeting's files.
4. **Write files** in `<ws>/<folder>/YYYY-MM-DD_slug/`. The slug is the title, lowercased and hyphenated, ASCII only, max 50 characters; add `-2`, `-3`… if a different meeting already uses that folder.
- `transcript.md`: a small header, then the transcript as received. Don't clean it up or summarize it. If it came from `_inbox/`, move the file's content here and delete the inbox file.
- `notes.md`: follow `<skill>/references/note-template.md` exactly.
- Upgrading a placeholder: write `transcript.md`, then rewrite `notes.md` as a full note.
- Upgrading a provisional note when a verbatim transcript arrives: if the `notes.md` file's SHA-256 still equals `notes_sha256` in its ledger line, the user hasn't edited it, so rewrite both files from the new transcript as `status: complete`. If the user has edited it, don't touch `notes.md`: save the new transcript as `transcript.md` (keeping the old one as `transcript.provisional.md`), set the ledger status to `complete`, and add a `_pending.md` item ("Full transcript arrived for <title>. Your edited notes were kept.").
5. **Ledger.** Add or replace this meeting's line in `.ledger.jsonl`, using the format in `<skill>/references/indexes.md`. Include `notes_sha256`, the SHA-256 of the `notes.md` you just wrote (`shasum -a 256` / `sha256sum`), so later runs can tell whether the user has edited it. The meeting counts as done only once its ledger line is written.
After all meetings:
6. **Cross-links.** Re-check `also_in` for the meetings filed in this run against the rules as they stand now, because a folder created later in the run may also match an earlier meeting. Update only front matter and ledger lines.
7. **Indexes.** Update `index.md`, the affected `_log/YYYY-MM.md` files and the affected `_project.md` files, following `<skill>/references/indexes.md`. Only touch folders and months that changed in this run. Never rewrite a generated file when only its "Updated" timestamp would change.
8. **Schedule.** Rewrite `<ws>/.schedule.json` (format in `<skill>/references/indexes.md`) with the **future** post-meeting checks (`at` later than now) for today's and tomorrow's attended meetings whose note is missing, `placeholder` or `provisional`. Only write the file if the list of checks changed; if the file doesn't exist yet, create it even when the list is empty. Leave out sensitive meetings. This file is what tells the trigger script when to start the next Quick run. Skip this step if no calendar is available.
9. **Leftover inbox files** older than the catch-up window that never matched: leave them, and report the count. Empty inbox subfolders stay where they are; they're the user's drop folders.
10. **Notify.** If anything changed and `notify` isn't `off`, send one desktop notification (≤150 characters, e.g. `4 meetings filed · 1 waiting on transcript · 6 new action items · 1 folder to review`), using the first of these that exists:
- macOS: `osascript -e 'display notification "<msg>" with title "Meeting notes"'`
- Linux: `notify-send "Meeting notes" "<msg>"`
- Windows: `powershell -NoProfile -Command "New-BurntToastNotification -Text 'Meeting notes','<msg>'"`; if that module is missing, skip it.
If none of them work, skip the notification silently. The summary still goes in the reply or run log.
11. **Reply** (on demand): the same summary, links to each new `notes.md`, any source errors, and the number of pending review items.
### Writing good notes
- Use only the transcript, the invite, and the notetaker's own notes. Never invent decisions, owners, dates or numbers. Anything unclear goes under Open questions.
- Every action item needs an owner: the attendee's name as spoken or as listed on the invite, or `@unassigned`. Only add a due date if one was said.
- Turn relative dates into absolute ones, counting from the **meeting date** ("by Friday" said on Wed Oct 28 → `due Fri Oct 30`). Keep a date as text if it can't be pinned to a day ("before next week's sync").
- If the source is a summary or dictation rather than a verbatim transcript (common for inbox files and invite attachments), say so in one short clause in the TL;DR.
- A task someone took on is an action item, even if it's also unresolved. Don't repeat it under Open questions unless there's a separate question nobody owns.
- Keep the TL;DR to 2–3 sentences, and group the Discussion by topic, not as a timeline.
- Write in the language the meeting was held in.
## Mode: Setup
The one-time routing proposal. Always on demand, and the user approves before anything is written. It works best with a calendar; without one, base the proposal on the transcripts in the ledger.
1. Gather the last `setup_lookback_days` of meetings (same filters as Run, sensitive meetings excluded).
2. Propose folders using `<skill>/references/routing.md` → "Proposing rules".
3. Show a table of each folder's name, its rules, how many meetings it would catch, and 2–3 example titles. Also list what would stay unsorted.
4. Let the user approve, rename, merge, split or drop folders.
5. On approval, write the rules to `routing.md` with `status: active` and create the folders, each with an empty `_project.md`.
6. Offer to backfill that period now (a Run over the range). Don't start it unasked.
## Mode: Review
Works through `_pending.md`. Always on demand.
1. **Pending folders** the agent created on its own: show each one with its rule and the meetings in it. The user can approve (`status: active`), rename, merge into another folder, or reject (the meetings are re-routed or go to `_unsorted/`).
2. **Before moving anything**, show a preview ("N meetings will move: A → B"). Move only after the user confirms, then update the ledger, every affected `_project.md`, the cross-links, `index.md` and the `_log` files.
3. Tick off the items you handled in `_pending.md`.
## Mode: Status
Report: the workspace path; which calendar and transcript sources were found (and which are missing); the last run (the newest `updated` in the ledger); the next post-meeting checks from `.schedule.json`; whether background triggers are running (the last few lines of `.trigger.log`, if it exists); how many items are in the inbox; how many placeholder notes there are; and open `_pending.md` items. Change nothing.
## Hard rules
- Never edit the body of a `notes.md` with `status: complete`; the user may have edited it. A `provisional` note is rewritten only if its hash shows it hasn't been edited. The only allowed change is its front matter (`folder`, `also_in`) during a confirmed move.
- Never write anything for a sensitive meeting.
- Never move filed meetings without a preview and confirmation.
- Never ask questions during an unattended run.
- Calendar and notetaker sources are read-only: never create, edit or delete events or recordings.
- Never delete a folder that still contains meetings.
- Never put personal data from meetings into anything outside `<ws>`: no URLs, no other files, no external services.
Pliki w paczce (10)
- SKILL.md 18 KB
- references/indexes.md 5 KB
- references/note-template.md 2 KB
- references/providers.md 9 KB
- references/routing.md 5 KB
- references/storage.md 5 KB
- templates/_pending.md 1 KB
- templates/config.md 2 KB
- templates/index.md 1 KB
- templates/routing.md 1 KB
Wersje
| Repozytorium | Wynik | Licencja |
|---|---|---|
| Ta strona · Andrew-eGov/meeting-notes-agent | 100 | MIT |
| ericrisco/rsc-harness | 100 | MIT |
| TerminalSkills/skills | 100 | Apache-2.0 |
| cdeweyx/conor-os | 100 | MIT |
| nimadorostkar/Claude-Skills-collection | 100 | MIT |
| SHOTClubhouse/shotos | 100 | MIT |
| microclaw/microclaw | 100 | MIT |
| scratchpad-sh/skills | 100 | MIT |
| codeape-7/ai-agent-workflowgroup | 100 | MIT |
Migawka bazy z 13.11.2026 · wersja paczki cb0f304964de (skrót SHA-256 całej paczki)
Pobrania w ostatnich 30 dniach: 0 · ranking
Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.
Podobne skille
- 100granola-setup
Łączy Granola przez oficjalne API do automatycznej synchronizacji spotkań i transkrypcji, także gdy spotkania się nie synchronizują. - 95process-meetings
Zamienia zsynchronizowane spotkania w zaktualizowane strony osób, wyciągnięte zadania i uporządkowane notatki, na przykład po synchronizacji z Granola lub Otter. - 96granola-sync
Importuje spotkania z Granola do folderu granola w Obsidianie jako notatki. - 100granola
Importuje, wylistowuje i eksportuje nagrania i transkrypcje spotkań z Granola przez Personal API, w tym eksport do notatek Obsidian w formacie zgodnym z Fathom. - 95granola-migration-deep-dive
Prowadzi przez migrację z Otter.ai, Fireflies, Fathom i podobnych narzędzi do notatek ze spotkań na Granola, wraz z planem przeniesienia danych. - 100meeting
Zamienia surową transkrypcję lub notatki ze spotkania, rozmowy albo 1:1 w uporządkowaną notatkę z kluczowymi punktami i zadaniami, domyślnie z folderu Inbox w Obsidian.