Meeting Summarizer

Meeting Summarizer

Audio → polished transcript → opinionated summary, in one Markdown file.

Transcription, diarization, polish, and speaker identification run via mac-meeting-transcriber (the mmt CLI) at /Users/KZFZ9H/Desktop/python_code/mac-meeting-transcriber. This skill orchestrates that pipeline and appends the summary. Questions are asked only when the LLM's content-based speaker inference is genuinely uncertain — never up front, "just to be safe".

When to apply

• User attaches an audio file (mp4, m4a, wav, mp3) and mentions meeting / summary / 总结.
• User types /meeting-summarizer.
• User says "summarize this meeting", "总结这个会议", or similar.
• User points at an existing transcript .md inside ~/Desktop/python_code/meeting_notes/.

Workflow

• 0. Resolve input (audio vs. existing transcript)
• 1. Identify "me" (default: Jason)
• 2. Idempotency check
• 3. Per-speaker key points
• 4. Intent analysis
• 5. Action items table
• 6. Append + verify

Step 0 — Resolve input

Case A: audio file

1. Resolve the audio path: explicit user path > Cursor attachment > most recently modified audio in ~/Downloads/.
2. Collect candidate speaker names — Jason plus unique names harvested from prior transcripts:

   rg '^- \*\*([^*]+)\*\*$' ~/Desktop/python_code/meeting_notes/*.md \
     -or '$1' --no-filename 2>/dev/null | sort -u
   

3. Run mmt via the Shell tool with working_directory: "/Users/KZFZ9H/Desktop/python_code/mac-meeting-transcriber" and block_until_ms ≥ 900000 (cold: ~12 min, warm cache-hit: ~3 min):

   uv run mmt "<audio_path>" --speakers "Jason,Leo,Kashish,..." -v
   

   Output lands in $MMT_OUTPUT_DIR (~/Desktop/python_code/meeting_notes/<stem>.md). The final stdout line is the full output path.
4. Scan -v logs for confidence=X.XX and mapping: {...}.
5. If confidence < 0.75 OR the mapping conflicts with obvious content cues, use AskQuestion to disambiguate. Include 1–2 sample utterances per ambiguous speaker. If the user corrects a name, patch the generated .md:
   • Update the ## Participants list.
   • Rewrite every **[MM:SS] OldName:** segment prefix.

Case B: transcript .md already referenced — skip 0.1–0.5 and jump to Step 1.

Case C: neither referenced — pick the most recently modified *.md in ~/Desktop/python_code/meeting_notes/. If it already has ## Action Items, confirm which meeting before proceeding.

Step 1 — Identify "me"

Default "me" = Jason unless Jason is not a participant. Ask only if genuinely ambiguous.

Step 2 — Idempotency

If the file already contains a summary section (## Summary, ## Intents, ## Action Items), do NOT overwrite silently — confirm with the user before re-summarizing.

Step 3 — Per-speaker key points

For each participant, extract 3–6 bullets: what they said, what they pushed for, what they flagged. Preserve their voice — do not rewrite into your own register.

Step 4 — Intent analysis

Surface 2–5 intents per participant:

| Participant | Intent | Evidence [MM:SS] |

"Intent" = what they actually want, not a literal paraphrase. Every row needs a timestamp and a short verbatim quote.

Step 5 — Action items

| Owner | Action | Due | Source [MM:SS] |

• Owner must be a named participant — never "me", "someone", "TBD".
• Include actions that are explicitly stated or strongly implied ("I'll send it over", "Leo will draft…").
• Due only when the transcript states it; otherwise leave blank.

Step 6 — Append + verify

Append ## Summary, ## Intents, ## Action Items after the existing transcript. Before returning, verify:

• Every participant appears in per-speaker points.
• Every Owner in actions is in ## Participants.
• All [MM:SS] citations point at real segments.

Language rules

• Preserve code-switching verbatim (中英混合 stays mixed). Never translate either direction.
• Quote key phrases inline rather than paraphrasing them.
• Keep Chinese quotes in Chinese, English quotes in English — even within one bullet.

Style

• Cite every factual claim about the meeting with [MM:SS].
• Separate facts from inference: prefix reading-between-the-lines content with Likely read: or similar.
• Prefer tight bullets over prose.

Anti-patterns

• Do NOT ask "who is SPEAKER_01?" up front if mmt's mapping is already confident.
• Do NOT flatten speaker voice or code-switching into neutral English.
• Do NOT invent action items the transcript doesn't support.
• Do NOT overwrite an existing summary without explicit confirmation.

Example

User attaches 4.15 Leo.mp4 and types /meeting-summarizer.

1. Resolve audio → collect speakers (Jason,Leo,Kashish,Rohan harvested from past notes) → run uv run mmt "4.15 Leo.mp4" --speakers "Jason,Leo,Kashish,Rohan" -v.
2. mmt logs: confidence=0.92 mapping: {'SPEAKER_00': 'Leo', 'SPEAKER_01': 'Jason'} → accept silently, no question asked.
3. Transcript saved to ~/Desktop/python_code/meeting_notes/4.15 Leo.md.
4. Append per-speaker points, intents table, and action items for Jason and Leo.
5. Return the path plus a one-line gist to the user.