--- name: "transcript-watcher" description: "Polls OneDrive Recordings/ for new Teams meeting MP4 files matching watched series name prefixes. On detection, invokes /meeting-transcript to download the .docx. Downstream /doc-watcher converts to .md and notifies. Detection-only; cheap polling, Playwright only when a new recording is found. Triggers: '/transcript-watcher', 'watch a meeting series for transcripts', 'detect new recordings'." --- # transcript-watcher Polls the user's OneDrive `Recordings/` folder for new Teams meeting MP4 files matching one or more watched meeting-series name prefixes. When a new MP4 is detected, hands off to the `/meeting-transcript` skill (Playwright) to download the .docx. The downstream `/doc-watcher` automation then converts the .docx to .md and sends the Teams notification. This skill is intentionally narrow: - **It detects new recordings.** It does NOT transcribe, summarize, or notify. - **Polling is cheap** (one Graph call per cycle). Playwright runs only when there is a new MP4 to extract. - **State** is the highest `lastModifiedDateTime` seen per watched series, so each MP4 fires extraction exactly once. ## Files - `config.json` — list of watches, recordings folder id, automation id - `SKILL.md` — this file ## Config shape `C:\Users\\.copilot\m-skills\transcript-watcher\config.json` ```json { "recordings_folder_id": "", "interval_minutes": 30, "automation_id": "", "watches": [ { "id": "my-meeting-series", "display_name": "My Meeting Series", "name_prefix": "My Meeting Series", "chat_id": "19:meeting_@thread.v2", "last_seen_modified": "2026-01-01T00:00:00Z" } ] } ``` - `recordings_folder_id` — OneDrive folder id for `Documents/Recordings`. Stable per user; reuse across watches. - `interval_minutes` — polling cadence (15 / 30 / 60). - `automation_id` — populated after `m_create_automation`. - `watches[].name_prefix` — the literal filename prefix Stream uses. Format is `-YYYYMMDD_HHMMSSUTC-Meeting Recording.mp4`. Match by `name.startswith(name_prefix + "-")`. - `watches[].chat_id` — meeting chat id; passed to `/meeting-transcript` so it knows which chat to open. - `watches[].last_seen_modified` — ISO 8601 UTC timestamp. Only files with `lastModifiedDateTime > last_seen_modified` are treated as new. ## When the user invokes `/transcript-watcher` Read `config.json` via `view`. If it does not exist → **first-run setup** (section A). Otherwise → **poll cycle** (section B). If the user explicitly says "dry run", "test", or "what would fire", run section B with the **dry-run flag** (do NOT invoke `/meeting-transcript`, do NOT update state — only report what would be triggered). ## A. First-run setup Collect via `m_ask_user`: 1. **Meeting display name** for this watch (e.g. "Weekly Standup"). Use this as the `name_prefix`. 2. **Polling interval** — 15 / 30 / 60 (default 30). Then auto-derive: 1. **Recordings folder id**: list OneDrive root via `m365_list_files(limit=50)` and find the entry named `Recordings`. If not found in the first page, page until found, or ask the user to provide a path. 2. **Chat id**: `m365_search_chats(query=display_name, limit=5)`. If exactly one result, use it. If multiple, surface them via `m_ask_user` and let the user pick. 3. **Baseline `last_seen_modified`**: list the Recordings folder (`m365_list_files(folderId=recordings_folder_id, orderBy="lastModifiedDateTime desc", limit=20)`), filter to items starting with `name_prefix + "-"`, take the max `modified`. If no matching files exist, set baseline to the current UTC time. This prevents history from re-firing. Write `config.json` via `create`. Then create the automation via `m_create_automation`: ``` name: transcript-watcher schedule: every minutes prompt: (the section B prompt verbatim — see "Automation prompt" at the bottom) enabled: false # important: leave OFF until the user explicitly enables it ``` Save the returned automation id back into `config.json`. Tell the user: "Watcher configured but NOT activated. Run `/transcript-watcher` to dry-run, or say `enable transcript watcher` to flip it on." ## B. Poll cycle For each `watch` in `config.json.watches`: 1. `m365_list_files(folderId=config.recordings_folder_id, orderBy="lastModifiedDateTime desc", limit=20)` 2. Filter the returned items in memory to those where: - `name` starts with `watch.name_prefix + "-"` - `mimeType` is `video/mp4` (or filename ends with `.mp4`) - `modified` (ISO timestamp) is strictly greater than `watch.last_seen_modified` 3. Sort the matches by `modified` ascending (so we process oldest-first; that way if extraction crashes mid-batch the state advance is monotonic). 4. For each new MP4: - **Dry run:** just report `{watch_id, filename, modified}` and continue. Do NOT invoke the transcript skill. Do NOT update state. - **Live run:** Invoke `/meeting-transcript` with the meeting **chat-id** and let it handle the recording-start-time lookup from the chat thread (see meeting-transcript SKILL.md Step 3). The meeting-transcript skill already names the output `--.docx` and drops it in the configured transcripts folder, which doc-watcher is already watching. We do not need to pass a filename — just the chat id. - After successful invocation, update `watch.last_seen_modified = file.modified` and persist `config.json` immediately (don't batch state updates — crash-safety). 5. If no matching new files for any watch: stay completely silent (no Teams ping, no chat output). The automation runs every 30 min and we don't want noise. ## C. Reconfigure - "add watch " → run setup steps 1 + auto-derive chat id + baseline for that watch, append to `watches[]`. - "remove watch " → drop that entry. - "enable transcript watcher" → `m_update_automation(id=automation_id, enabled=true)`. - "disable transcript watcher" → `m_update_automation(id=automation_id, enabled=false)`. - "change interval to " → update config + `m_update_automation(schedule="every minutes")`. ## D. Failure modes (handle defensively) - `/meeting-transcript` extraction fails (Stream not ready, Playwright timeout, transcript not yet processed) → do NOT advance `last_seen_modified`. Report the error in chat (or via Teams DM if running from automation) so the user knows. Next poll will retry. - Two new MP4s in a single meeting (Teams sometimes splits recordings on reconnect) → process each one. The chat-thread "Meeting started at ..." will identify each occurrence's start time correctly, so filenames won't collide. - Multiple watches share `name_prefix` collisions (e.g. "Sync" matches several series) → require name prefixes that are unique. The skill does not de-collide; the user must pick distinct prefixes. ## Automation prompt When creating the automation in section A, pass this as the prompt: ``` Run /transcript-watcher in poll-cycle mode. Do not prompt the user. If any new MP4 is found, invoke /meeting-transcript with the watch's chat_id for each new file, then persist the updated last_seen_modified to C:\Users\\.copilot\m-skills\transcript-watcher\config.json. If no new files, stay completely silent — no Teams message, no chat output. If /meeting-transcript fails for any file, do NOT advance last_seen_modified for that file and send a single Teams DM via m_send_teams_message: "transcript-watcher: extraction failed for (). Will retry next cycle. Sent on your behalf by Clawpilot 🤖" ``` ## Privacy Meeting recordings + transcripts are sensitive. This skill never sends transcript content to chat or Teams; it only triggers the existing downstream skills, which already follow the user's notification preferences.