meeting-transcript

name	meeting-transcript
description	Fetch a meeting transcript — full, paginated, or filtered to a single speaker or phrase — without leaking raw PII into long-term storage.
category	meeting-notes
version	1
compatibility	devboy-tools >= 0.18
activation	["transcript of meeting","what did Alice say about","show the meeting transcript","quote from the meeting","what was said about"]
tools	["get_meeting_transcript","get_meeting_notes"]

meeting-transcript

Retrieve the speaker-attributed, timestamped transcript for a single meeting. Transcripts are routinely large (thousands of sentences for a one-hour call), so the skill is built around pulling only what the question actually needs rather than the whole thing.

When to use

• The user asks for a transcript by meeting title, id, or relative reference ("the call yesterday", "the one with Alice on Tuesday").
• The user wants a specific quote ("what did Bob say about the migration?").
• A follow-up skill (meeting-to-tasks) needs the raw text to extract action items that are not already in action_items.

If the user does not yet know which meeting — route through meeting-search first.

Procedure

1. Resolve the meeting id

You need a meeting_id. If the user already pasted one, skip this step. Otherwise confirm the id via get_meeting_notes (or via meeting-search for a keyword lookup) and pin the id before fetching:

devboy tools call get_meeting_notes '{
  "from_date": "2026-04-16T00:00:00Z",
  "to_date":   "2026-04-17T00:00:00Z",
  "limit": 10
}'

Confirm the title and date with the user before proceeding — pulling the wrong transcript is an irreversible PII leak into the chat.

2. Fetch the transcript

devboy tools call get_meeting_transcript '{"meeting_id": "<id>"}'

devboy tools call takes a tool name + a JSON argument object — it does not accept --budget (or any other CLI flag) for the tool itself. The format-pipeline budget is applied inside the tool runtime. The transcript output is rendered as plain text, not JSON; a long transcript may be shortened by the pipeline and picked up again via a subsequent call.

3. If the rendered transcript is incomplete, fetch the next chunk

If the response ends mid-thought, pull the next segment:

devboy tools call get_meeting_transcript '{"meeting_id": "<id>", "chunk": 2}'

chunk is a per-call argument understood by the transcript tool (the default is 1). The response is still text — do not expect a JSON body with fields like total_chunks or chunk_number. The tell for "more to fetch" is that the text ends mid-sentence, not a structured field.

4. Prefer a targeted fetch over a full dump

When the question is specific ("what did Alice say about X?"), do not scroll the whole transcript in the chat. Two patterns:

1. Search first, transcript second. Run meeting-search with the topic word, confirm this is the right meeting, then open the relevant chunk.

2. Grep the chunk. The transcript is rendered as text lines shaped like [mm:ss] <speaker>: <sentence>, so filter with grep / ripgrep:

   devboy tools call get_meeting_transcript '{"meeting_id": "<id>"}' \
     | grep -F '] Alice: '
   

5. Quote, do not paste

When you respond to the user, reply with the minimum quote that answers the question — typically one to three sentences plus the speaker and timestamp. Never paste the full transcript into the chat unless the user explicitly asked for it.

Success criteria

• The user gets an answer grounded in the transcript, with speaker + timestamp, in fewer than 10 lines for a targeted question.
• For a "give me the whole transcript" request, the skill streams chunk 1, explains how to request chunk 2+ if the user wants more, and waits for the user to ask rather than pre-fetching all chunks. (The pipeline does not surface a total-chunk count, so the skill must not promise one.)
• No untrimmed transcript content is copy-pasted into files the user did not name.

Guardrails

• PII. Transcripts contain names, emails, internal project names, and sometimes customer data. Do not write them to long-term storage (new docs, new issues, commit messages) unless the user explicitly asks for a specific excerpt. Summarise; do not mirror.
• Do not translate or edit quotes. If the user asks "what did Alice say?", answer with her words, not a paraphrase — downstream uses (legal, HR, customer conversations) rely on the exact phrasing.
• Confirm the meeting before fetching. A wrong meeting_id leaks a different meeting's contents into the chat and is not retractable.

Non-goals

• Editing or redacting transcripts. The skill is read-only.
• Summarising a full meeting. Use the summary and action_items fields returned by get_meeting_notes — those are provider-generated and cheaper to fetch than the full transcript.
• Cross-meeting aggregation. One transcript at a time.